ConvMAE: Masked Convolution Meets Masked

Autoencoders

Peng Gao1 Teli Ma1 Hongsheng Li1,2 Ziyi Lin2 Jifeng Dai3 Yu Qiao1
1
Shanghai AI Laboratory 2 MMLab, CUHK
3
SenseTime Research
arXiv:2205.03892v2 [cs.CV] 19 May 2022

Abstract

Vision Transformers (ViT) become widely-adopted architectures for various vision

tasks. Masked auto-encoding [2, 1, 28, 55] for feature pretraining and multi-
scale hybrid convolution-transformer architectures [12, 21, 49, 34, 57] can further
unleash the potentials of ViT, leading to state-of-the-art performances on image
classification, detection and semantic segmentation. In this paper, our ConvMAE
framework demonstrates that multi-scale hybrid convolution-transformer can learn
more discriminative representations via the mask auto-encoding scheme. However,
directly using the original masking strategy leads to the heavy computational cost
and pretraining-finetuning discrepancy. To tackle the issue, we adopt the masked
convolution to prevent information leakage in the convolution blocks. A simple
block-wise masking strategy is proposed to ensure computational efficiency. We
also propose to more directly supervise the multi-scale features of the encoder to
boost multi-scale features. ConvMAE-Base improves ImageNet-1K finetuning
accuracy by 1.4% compared with MAE-Base. On object detection, ConvMAE-
Base finetuned for only 25 epochs surpasses MAE-Base fined-tuned for 100 epochs
by 2.9% AP box and 2.2% AP mask respectively. Code and pretrained models are
available at https://github.com/Alpha-VL/ConvMAE.

1 Introduction
Self-supervised learning frameworks, such as DINO [6], MOCO-V3 [10], MAE [28], unleash
the potential of Vision Transformers (ViT) and achieve high performance on various downstream
vision tasks [33, 30, 58]. Among them, Mask Autoencoders (MAE) [28] demonstrate superior
learning ability and scalability. Motivated by BERT [15, 46, 4] in natural language processing,
MAE utilizes an asymmetric encoder and decoder architecture, in which masked tokens of the
encoder are reconstructed by the decoder. Experiments show that MAE can learn discriminative
and scalable representations from ImageNet-1K [14] without relying on large-scale datasets, such as
ImageNet-22K.
Local inductive bias [49, 21, 34, 12, 19, 57] and hierarchical representations [42, 53] are explored
for boosting the performance of ViT. The combination of local convolution and global transformer
operations leads to clear improvements on image classification [33], object detection [30], and
semantic segmentation [58]. In contrast to MAE [28], well-performing multi-scale backbones built
upon local and global operations are mainly trained in supervised manner. A natural question is
whether multi-scale backbone with local and global operations, which show promising performance on
supervised learning can be exploited to enhance the masked auto-encoding paradigm [28, 15, 2, 65].
In this paper, a simple and effective self-supervised learning framework, dubbed as ConvMAE, is
proposed to train scalable representations by introducing hybrid convolution-transformer architectures
and masked convolution into the masked auto-encoders. Although the modifications to the original
MAE are minimal, ConvMAE shows great success on pretraining visual representations for boosting
the performances of various tasks.

Preprint. Under review.

Different from MAE [28], the encoder of ConvMAE progressively abstracts the input image into
multi-scale token embedding, while the decoder reconstructs the pixels corresponding to masked
tokens. For high-resolution token embedding at early stages, convolutions blocks are adopted to
encode local content. For low-resolution token embedding at late stage, transformer blocks are used
to aggregate global context. The encoder therefore obtains both local and global FOV at different
stages and generates discriminative multi-scale features. Note that the ConvMAE encoder is partly
motivated by the strong hybrid convolution and transformer backbones, including Co-AtNet [12],
Early Convolution [57], Container [21] and Uniformer [34]. However, previous hybrid convolution-
transformer networks were either not explored for masked auto-encoding [21, 34, 20] or show
very similar performance to MAE [52, 59]. Instead of designing novel architectures, we focus on
making basic hybrid convolution-transformer architectures work for mask auto-encoding and conduct
extensive experiments to demonstrate its effectiveness on various downstream tasks.
The efficient and effective training of ConvMAE is enabled by a block-wise masking strategy
with masked convolution [60, 25, 31, 48, 24, 40]. The masking strategy adopted in current mask-
autoencoding frameworks, such as BEiT [2], MAE [28], SimMIM [59], cannot be naively used for
ConvMAE as all tokens need to be kept in the later transformer stages. This leads to unaffordable
computation cost for pretraining large and huge models, losing MAE’s efficiency advantage of
omitting masked tokens in transformer encoder. In addition, directly pretraining with the convolution-
transformer encoder causes pretraing-finetuning discrepancy as only visible tokens are processed
during finetuning stages.
To tackle the issues, we focus on designing hybrid convolution-transformer architectures suitable
for mask auto-encoding. Specifically, our ConvMAE adopts a block-wise masking strategy to
first obtain a mask for the late stage in transformer and then progressively upsamples the mask to
larger resolutions in early convolutional stages. In this way, tokens processed by late stages can be
completely separated into masked tokens and visible tokens and inherit the computation efficiency
of MAE. To prevent information leakage, the convolution blocks at early stages are equipped with
masked convolutions, which avoid mixing up features of masked and visible regions in late stages
to ensue the training effectiveness. Masked convolution has been well explored in sparse feature
extraction [25, 48, 24, 60] and image inpainting [40]. It can be naturally integrated into the hybrid
convolution-transformer architecture to enable masked auto-encoding.
Our ConvMAE can naturally provide multi-scale features for object detection and semantic segmen-
tation, which are required by modern detection [30] and segmentation frameworks [58]. Multi-scale
features from the pretrained ConvMAE can significantly improve the performances of object detection
and semantic segmentation compared with MAE. ConvMAE with masked-based autoencoding can
even surpass the fully-supervised pretraining of Swin and MViT [42, 36].
In summary, our contributions can be summarized below: (1) We present the strong and efficient self-
supervised framework ConvMAE, which is easy to implement but show outstanding performances
on different tasks. (2) The proposed ConvMAE naturally generates hierarchical representations and
exhibit promising performances on object detection. (3) ConvMAE-Base improves the ImageNet
finetuning accuracy by 1.4% compared with MAE-Base. On COCO 2017 with Mask-RCNN,
ConvMAE-Base achieves 53.2% AP box and 47.1% AP mask with a 25-epoch training schedule while
MAE-Base attains 50.3% AP box and 44.9% AP mask with 100 training epochs. On ADE20K with
UperNet, ConvMAE-Base surpasses MAE-Base by 3.6 mIoU (48.1% vs. 51.7%).

2 Approach
2.1 A Brief Revisit of MAE
Masked Autoencoders (MAE) [28] is a self-supervised method for pretraining ViT by reconstructing
masked RGB patches from visible patches. Although MAE has a simple design, it has been proven
to be a strong and scalable pretraining framework for learning visual presentations. MAE consists
of transformer-based encoder and decoder, where only visible patches are fed into the encoder
and learnable mask tokens are processed by the decoder for image reconstruction to learn visual
representations. As the encoder only needs to process a small portion of visible tokens, it alleviates
the scalability problem to pretrain large vision models.
2.2 ConvMAE
ConvMAE is a simple and effective derivative of the popular MAE [28] with minimal but effective
modifications on the encoder design and the masking strategy. The goal of ConvMAE is to learn

2
 H×W×3
W/4 Block-wise Masking
W/8
W/16

UpSample UpSample
H/4 H/8 H/16

Patch Embedding
Patch Embedding

Patch Embedding

Transformer
Convolution
Convolution

Transformer
Masked

Masked

Block

Linear
Block

Block

Block
×2 ×2 ×11
Stage1 Stage2 Stage3
H/4×W/4×C1 H/8×W/8×C2 Encoder (H/16×W/16)×C3 Decoder

H/4×W/4×C1 H/8×W/8×C2 Masked Convolution Block

Feature Embeddings

DepthWise Conv
StrideConv+Flatten StrideConv+Flatten Mask

Linear

FFN
Mask Mask Value

Multi-Scale Fusion

Figure 1: The pipeline of our proposed ConvMAE which consists of a hybrid convolution-transformer
encoder, block-wise masking strategy with masked convolution and multi-scale decoder.

discriminative multi-scale visual representations and to prevent pretraining-finetuning discrepancy

when applies MAE [28] on convolution-transformer networks.
Directly applying the original masking strategy on the feature maps of the convolution-transformer
encoder would make transformer layers keeping all tokens during the pretraining, jeopardizing the
training efficiency. We introduce a hierarchical masking strategy coupled with masked convolution for
the convolution stages to ensure only a small number of visible tokens are input into the transformer
layers. The overall pipeline of ConvMAE is shown in Figure 1.
The Hybrid Convolution-transformer Encoder. There are previous strong hybrid convolution-
transformer architectures, such as Co-AtNet [12], Container [21], BoTNet [49], Uniformer [34] and
Early Conv [57]. Without using such complicated architectures, we show that a simple design of
multi-scale convolution-transformer encoder can already learn powerful representations for various
downstream tasks. As shown in Figure 1, our encoder consists of 3 stages with output spatial
resolutions of H W H W H W
4 × 4 , 8 × 8 , 16 × 16 , respectively, where H × W is the input image resolution.
The first two convolutional stages use convolution blocks to transform the inputs to token embeddings
H W H W
E1 ∈ R 4 × 4 ×C1 and E2 ∈ R 8 × 8 ×C2 . Our convolution blocks follow the design principle of the
transformer block by only replacing the self-attention operation with the 5 × 5 depthwise convolution
The third transformer stage uses commonly used self-attention blocks to obtain token embeddings
H W
E3 ∈ R 16 × 16 ×C3 . Between every stage, stride-2 convolutions are used to downsample the tokens
to half of its previous spatial resolution. The local convolutions in stages 1 and 2 have relatively
small field-of-view, the transformer blocks in stage 3 aggregate and fuse features from the coarse-
grained features and extend the field of view to the whole image. Different from other ViTs, such as
CPT [11], Container [21], Uniformer [34], CMT [26], Swin [42], which replace absolute position
embedding [42] with relative position embedding or zero-padded convolution at the inputs of the
first stage [11, 21, 34, 26], we find that adding absolute position embeddings to the inputs of the
transformer stage-3 leads to the optimal performance. The class token is also removed from our
encoder which shows limited influence.
Block-wise Masking with Masked Convolutions. Mask auto-encoders, such as MAE [28] and
BEiT [2], adopt a random mask on the input tokens. However, the same strategy cannot be directly
applied to our ConvMAE encoder. Uniformly masking stage-1 input tokens from the H4 × W 4 feature
maps would cause all tokens of stage-3 to have partially visible information and requires keeping all
stage-3 tokens. Therefore, we propose to first generate the random mask to mask out p% (e.g., 75%) of

3
 Stage1 Stage2 Stage3 MS-Encoder

Transformer Block

Patch Embedding

Patch Embedding

Patch Embedding
H×W×3 Convolution Convolution Down
Block Block Sample

Local or Global Attention

×2 ×2 ×11

MaskRCNN/UperNet E1 E2 E3 E4
H/16×W/16×C3 H/32×W/32×C4
H/8×W/8×C2
H/4×W/4×C1

Figure 2: Overview of finetuning ConvMAE for object detection and semantic segmentation. The
intermediate features of different stages serve as multi-scale inputs for an FPN [38] module.

stage-3 input tokens and upsample the mask by 2 times and 4 times to obtain the corresponding block-
wise masks for masking stage-2 and stage-1 inputs, respectively. The corresponding masked tokens in
the three stages are dropped in the encoding process and are reconstructed by the decoder for feature
learning. In this way, ConvMAE only needs to keep as few as 25% tokens in the time-consuming
transformer blocks for training and the efficiency of ConvMAE is not compromised.
However, the 5 × 5 depthwise convolutions in the first two stages naturally lead to receptive fields
larger than the masked patches and cause information leakage when reconstructing masked tokens. To
avoid such information leakage and ensure the quality of pretraining, we adopt masked convolution
[25, 48] in the first two stages, so that the masked regions would never be involved in the encoding
process. The use of masked convolution is crucial to the superior performance of ConvMAE and the
pretraining-testing discrepancy is prevented by removing partially masked tokens from stage.
The Multi-scale Decoder and Loss. The decoder of the original MAE [28] takes as input both
visible tokens Ed from the encoder and the mask tokens [Mask], and transform them in stacked
transformer blocks for image reconstruction. Our ConvMAE encoder obtains multi-scale features E1 ,
E2 , E3 , captures both fine- and coarse-grained image information. To better supervise the pretraining
of such multi-grained representations, we downsample E1 and E2 to the same size of E3 with stride-4
and stride-2 convolutions and fuse multi-grained tokens via a linear layer to obtain visible tokens Ed ,

Ed = Linear(StrideConv(E1 , 4) + StrideConv(E2 , 2) + E3 ), (1)

where StrideConv(·, k) represents stride-k convolution. The multi-scale decoder is illustrated in the
bottom-left part of Figure 1. The same losses from MAE [28] are used for reconstructing masked
image patches and only the reconstruction of masked patches are considered in the objective function.

2.3 ConvMAE for Object Detection and Semantic Segmentation

After pretraining, the proposed ConvMAE can naturally generate multi-scale feature maps, which
can be processed by existing object detection and semantic segmentation heads.
As shown in Figure 2, to finetune ConvMAE for object detection, an E4 feature map of 1/32 input
resolution is first obtained by 2 × 2 max pooling E3 . However, as the ConvMAE stage-3 has 11 global
self-attention layers (in our ConvMAE-base model) with excessive computational cost, we follow
Benchmarking ViT [37] to replace all but 1st, 4th, 7th, 11th global self-attention layers in stage-3
to shifted-window local self-attention layers [42] with alternatively shifted 7 × 7 windows. The
modified local self-attention layers are still initialized by the pretrained global self-attention layers. A
global relative position bias [2, 42, 28, 37] is shared between global transformer blocks. Similarly,
a local relative position bias [2, 42, 28, 37] is shared by local transformer blocks. In this way, the
heavy computational and GPU memory costs of the stage-3 are much mitigated. The multi-scale
features E1 , E2 , E3 , E4 are then fed into the MaskRCNN [30] head for object detection. To finetune
ConvMAE for semantic segmentation, its stage-3 architecture is kept as the images in segmentation
datasets have relatively smaller resolutions. The multi-scale features are feed into UperNet [58].

4
2.4 ConvMAE for Video Understanding
Attention based models [54, 63, 50, 3, 42] have demonstrated superior performance on video under-
standing. Our ConvMAE can also be extended to serve as a strong video pretraining framework,
dubbbed as VideoConvMAE, with simple modifications. Specifically, VideoConvMAE replaces
image patch embedding with cube embedding, after which stage 1 and stage 2 perform local spatial-
temporal feature fusion with masked 3D convolutions. Stage 3 still adopts stacked transformer
blocks for spatial-temporal fusion. The spatial position embedding is extended to spatial-temporal
embedding. Similar to the multi-scale decoder proposed in Section 2.2, outputs from stages 1, 2 and 3
are fused before feeding into a spatial-temporal transformer decoder for masked pixel reconstruction.
Details about VideoConvMAE pretraining are in appendix B. Note that unlike previous approaches,
which initialize models pretrained on images [34, 36, 7], our VideoConvMAE is pretrained from
scratch on pure video datasets.
3 Experiments
To validate our proposed ConvMAE, we conduct experiments of image classification on ImageNet-
1K [14] dataset. The pretrained ConvMAE is also extensively tested on object detection and semantic
segmentation. By default, we report performance of our the ConvMAE-base model with multi-scale
decoder, which has similar parameters and FLOPs as the MAE-base.
3.1 ImageNet-1K Pretraining and Finetuning
Experimental Setup. ImageNet-1K [14] consists of 1.3M images of 1k categories for image
classification and is split to the training and validation sets. We pretrain our ConvMAE on ImageNet-
1K training set. By default, we fix the mask ratio to 25% following the original MAE [28]. The
decoder is designed to have 8 transformer layers with 512 feature dimensions and 12 attention heads.
We adopt a 1600-epoch cosine learning rate schedule with the first 40 epochs for warming up. The
AdamW optimizer is utilized with a base learning rate of 1.5 × 10−4 , a weight decay of 0.05 and a
batch size of 1024. Random cropping is employed as data augmentation during pretraining. After
pretraining, the ConvMAE encoder is used for supervised finetuning on ImageNet-1K training set for
100 epochs using the cosine learning rate schedule. We follow the default finetuning parameters of
the original MAE [28] except for the layer-wise learning-rate decay parameters (0.65, 0.75, 0.85).
For finetuning, we report the classification accuracy on the ImageNet validation set of the finetuned
and pretrained (linear probe) ConvMAE encoders.
Results on ImageNet-1K Finetuning. We report the accuracy of ConvMAE on Table 1 and conduct

Methods Backbone Params. (M) Supervision Encoder P-Epochs FT (%) LIN (%)
BEiT [2] ViT-B 88 DALLE 100% 300 83.0 37.6
MAE [28] ViT-B 88 RGB 25% 1600 83.6 67.8
SimMIM [59] Swin-B 88 RGB 100% 800 84.0 N/A
MaskFeat [55] ViT-B 88 HOG 100% 300 83.6 N/A
data2vec [1] ViT-B 88 Momentum 100% 800 84.2 N/A
ConvMAE ConViT-B 88 RGB 25% 1600 85.0 70.9
Table 1: Comparison with state-of-the art mask auto-encoding schemes with similar model size. FT
and LIN denotes ImageNet-1K finetuning and linear probe accuracy respectively.

comparisons with state-of-the-art mask autoencoding methods. BEiT [2] pretrains ViT-B through the
prediction of visual tokens tokenized by the DALL-E encoder. With 300-epoch pretraining, BEiT
can reach a finetuning accuracy of 83.0% and a linear-probe accuracy of 37.6%. Compared with
BEiT, ConvMAE processes only 25% visible tokens in the encoder and has a lightweight decoder
for reconstruction. ConvMAE can surpass its finetuning accuracy and linear-probe accuracy by
large margins (+2.0%/+33.3%). Compared with the original MAE pretrained for 1,600 epochs,
our ConvMAE surpasses its finetuning accuracy by 1.4% with same number of pretraining epochs.
SimMIM [59] adopts a Swin-B [42] to generate hierarchical representations. ConvMAE achieves
improvement over its finetuning accuracy (+1.0%). MaskFeat [55] uses HOG [13] features as
prediction targets. Data2vec [1] incorporates a momentum encoder [29] to generate predictions in an
online manner. Both MaskFeat and Data2vec have higher computational costs than our ConvMAE.
They can be considered as complementary directions for improving the mask auto-encoding scheme.
3.2 Object Detection
Experimental Setup. COCO dataset [39] has been widely adopted for benchmarking object detection
frameworks. Mask-RCNN [30] is one of the most popular frameworks for object detection. We

5
 Methods Pretraining P-Epochs F-Epochs AP box AP mask Params (M) FLOPs (T)
Benmarking [37] IN1K w/o labels 1600 100 50.3 44.9 118 0.9
ViTDet [35] IN1K w/o labels 1600 100 51.2 45.5 111 0.8
MIMDET [20] IN1K w/o labels 1600 36 51.5 46.0 127 1.1
Swin+ [42] IN1K w/ labels 300 36 49.2 43.5 107 0.7
MViTv2 [36] IN1K w/ labels 300 36 51.0 45.7 71 0.6
ConvMAE IN1K w/o labels 1600 25 53.2 47.1 104 0.9
Table 2: Performances of different pretrained backbones on object detection with Mask-RCNN [30].
Models Pretrain Data P-Epochs mIoU Params (M) FLOPs (T)
DeiT-B [51] IN1K w/ labels 300 45.6 163 0.6
Swin-B [42] IN1K w/ labels 300 48.1 121 0.3
MoCo V3 [29] IN1K 300 47.3 163 0.6
DINO [6] IN1K 400 47.2 163 0.6
BEiT [2] IN1K+DALLE 1600 47.1 163 0.6
PeCo [17] IN1K 300 46.7 163 0.6
CAE [9] IN1K+DALLE 800 48.8 163 0.6
MAE [28] IN1K 1600 48.1 163 0.6
ConvMAE IN1K 1600 51.7 153 0.6
Table 3: Comparison with different pretrained backbones on ADE20k with UperNet.

employ the encoder of pretraied ConvMAE as the backbone for Mask-RCNN. We finetune Mask-
RCNN on COCO train2017 split and report AP box and AP mask on val2017 split. We follow most
setups of Benchmarking ViT [37]. We report the model performance on object detection under a 25
epochs cosine schedule with a base learning rate of 8.0 × 10−5 , a weight decay of 0.1.
Results on COCO 2017. We compare the performances of state-of-the-art visual backbones in
Table 2. Benchmarking ViT [37] extensively explores using plain ViT with Mask-RCNN. Compared
with Benchmarkin ViT [37] finetuned for 100 epochs on COCO, ConvMAE can significantly improve
AP box and AP mask by 2.9% and 2.2% with 25 finetuning epochs. ViTDet [35] improves Bench-
marking ViT [37] by introducing a simple feature pyramid module. MIMDet [20] adds a randomly
initialized convolution stem and randomly drops input tokens to increase the training efficiency of
Mask-RCNN [30]. Note that MIMDet [20] introduces extra parameters due to the incorporation of
MAE decoder. Compared with improved version of Benchmarking ViT [37], such as ViTDet [35]
and MIMDet [20], ConvMAE surpass them by 2.0% and 1.7% with a shorter finetuning schedule
(25 epochs vs 100/36 epochs), fewer parameters (104M vs 111M/127M) and similar FLOPs (0.9T).
This validates the effectiveness of our proposed ConvMAE framework. Swin [42] and MViTv2 [36]
are state-of-the-art hierarchical visual backbones. Although adopting a simpler multi-stage archi-
tecture, ConvMAE outperforms Swin annd MViTv2 by 4.0%/3.6% and 2.2%/1.4% in terms of
AP box /AP mask . Note that Swin [42] and MViT [36] v2 are pretrained for 300 epochs with 100%
tokens in a supervised manner while ConvMAE is only pretrained using masked autoencoder with
25% visible tokens, which is efficient for object detection.
3.3 Semantic Segmentation
Experimental Setup. ADE20K [64] is a widely-used semantic segmentation dataset which contains
25,562 images of 150 fine-grained categories. The dataset is split into training, validation, and testing
sets. We leverage the UperNet [58], a hierarchical segmentation network head to compare ConvMAE
with other backbones. Our ConvMAE with UperNet [58] is finetuned on ADE20K training set and
tested on validation split. In the training phase, the backbone is initialized with the weights pretrained
for 1600 epochs on ImageNet-1K and other modules are initialized with Xavier initialization. We
adopt a 16k-iteration polynomial learning rate schedule with the first 1500 iterations for warming up.
The AdamW [44] optimizer is adopted with an initial learning rate of 10−4 , a weight decay of 0.05
and a batch size of 16. We follow the default finetuning configurations of MAE on ADE20K except
for the feature dimensions for the decoder head and the layer-wise learning rate decay is set as 0.75.
Results on ADE-20K. We report the Mean Intersection over Union (mIoU) performance of Con-
vMAE and other state-of-the-art backbones in Table 3. With the 300-epoch pretraining, MoCo
V3 [10] can reach 47.2 mIoU when finetuned on semantic segmentation. BEiT [2], PeCo [17]
and CAE [9] utilize discrete VAE as visual tokenizer to create the targets. Both BEiT and CAE
adopt the DALL-E [47] codebook trained on 250M images, while PeCo trains a codebook only

6
 82.7 70.7

81.7 69.9 70.3

70.0
80.1 69.3
79.8 67.7
79.4 67.6
78.8

77.8 66.1

Figure 3: Finetuning accuracy on Kinetics-400 and Something-Something-v2.

on ImageNet-1K. Compared with these methods, our 1600-epoch pretrained ConvMAE achieves
much higher performance (51.7%). Compared with MAE pretrained 1600 epochs, our ConvMAE
outperforms it by 3.6% mIoU, demonstrating the hierarchical representations of ConvMAE largely
diminishes the transfer gap between pretrained backbones and downstream networks.
3.4 Video Understanding
Experimental Setup. To validate the video understanding ability of VideoConvMAE, we pretrain
on Kinetics-400 (K400) [32] and Something-something V2 (SSV2) [23] independently and report
the finetuning accuracy on K400 and SSV2. The video pretraining and finetuning protocol closely
follow the image protocol explained in Section 3.1. For SSV2, we finetune 50 epochs and turn off
the flip augmentation. Unlike random masking in image pretraining, tube masking [50] with 90%
mask ratio proposed by VideoMAE is adopted as the default masking strategy. For testing, we use
the same number of views as VideoMAE for fair comparison, i.e., 3 spatial × 5 temporal views for
Kinetics-400 and 3 spatial × 2 temporal views for Something-Something-v2. All results are reported
using only the finetuning dataset without extra image or video data.
Results on K400 and SSV2. We compare the finetuned accuracy on K400 and SSV2 with Video-
MAE [50] for different pretraining epochs. As shown in Figure 3, VideoConvMAE outperforms
VideoMAE by a clear margin at 200 and 800 pretraining epochs. Notably, on Kinetics-400, Video-
ConvMAE pretrained for 200 epochs slightly outperforms VideoMAE at 1600 epochs, and 800-epoch
pretrained VideoConvMAE with multi-scale decoder outperforms VideoMAE at 1600 epochs by
more than 2.9%. On Something-Something-v2, our 800-epoch model with multi-scale decoder
slightly outperforms VideoMAE at 2400 epochs, which indicates 3x reduction in pretraining epochs.

ImageNet COCO ADE20K

Pretrain Epochs
FT LIN AP box AP mask mIoU
200 84.1 62.5 50.2 44.8 48.1
400 84.4 66.9 51.4 45.7 49.5
800 84.6 68.4 52.0 46.3 50.2
1600 84.6 69.4 52.5 46.5 50.7
Table 4: The influence of increasing pretraining epochs on various downstream tasks.

3.5 Ablation Study of ConvMAE

We conduct extensive ablation studies on ConvMAE to analyze different components of ConvMAE
(see Table 5 and 6). By default, we report the performance of ConvMAE without multi-scale decoder
during ablation studies.
Pretraining epochs. For MAE, longer pretraining epochs can significantly improve the learned
representations learned. We pretrain ConvMAE-Base with 200, 400, 800 and 1600 epochs to test
the influences on ConvMAE. We report the ImageNet-1K finetuning (FT) and linear probe (LIN)
accuracies, AP box and AP mask of COCO, mIoU of ADE20K on Table 4. We observe improved
performances on most downstream tasks with longer pretraining epochs.
Input token random masking. As shown in Table 5, we replace the proposed block-wise mask
strategy with MAE’s input token random masking. Compared with our ConvMAE-base, the ImageNet-
1K finetuning accuracy drops from 84.6% to 84.2% which validates that the proposed simple block-
wise masking strategy can alleviate pretraining-finetuning discrepancy. Input-token random masking

7
results in all tokens in stage-3 being processed by computationally intensive transformer blocks and
causes FLOPs to increase by 1.7×.
Influence of masked convolution. Masked Convolution can prevent information leakage due to
the overlapping window in convolution. Removing masked convolution decreases the ImageNet-
1K finetuning accuracy from 84.6% to 81.5% , which demonstrates that information leakage in
convolution stages hinders feature learning in mask autoencoding.
Convolution kernel sizes in stages 1 and 2. Enlarging the kernel size in convolution is shown to be
effective for semantic segmentation [45] and visual backbone designs [16, 43]. We also test with
enlarging the 5 × 5 kernel size in stages 1 and 2 to 7 × 7 and 9 × 9. As shown by Table 5, we observe
that larger kernel sizes barely influence the performance of ConvMAE on ImageNet-1K accuracy.
We hypothesize that the transformer blocks in stage-3 already provide a global FOV which can cancel
out the gains introduced from large kernels.
Multi-scale Decoder. In Table 6, we incorporate multi-scale decoder into ConvMAE-base and
pretrain for 200 and 1600 epochs. Comparared with ConvMAE pretrained 200 epochs, multi-scale
decoder can improve classification accuracy, detection AP box , detection AP mask and segmentation
mIoU by 0.3%, 0.6% 0.6% and 0.4%, respectively. Given longer pretraining, multi-scale decoder
can improve classification accuracy, linear probe accuracy, detection AP box , detection AP mask and
segmentation mIoU by 0.4%, 1.6%, 0.7%, 0.6%, 1.0%, respectively. This indicates that fusing
multi-grained tokens for mask reconstruction can lead to powerful representations. We will explore
more advanced multi-scale decoder architectures such as UNet in the future.

84.6 69.4 52.4

83.7

50.4
67.8

58 38
16

(a) ImageNet Finetuning (b) ImageNet Linear Probing (c) COCO Detection

Figure 4: Convergence of MAE and ConvMAE on various tasks.

Masked Block 5×5 7×7 9×9

P-Epochs FT (%) FLOPs
Conv Masking Conv Conv Conv
3 3 3 7 7 84.6 1×
3 7 3 7 7 84.2 1.7×
7 3 3 7 7 81.5 1×
800
3 3 3 7 7 84.5 0.997×
3 3 7 3 7 84.4 1.003×
3 3 7 7 3 84.6 1.007×
Table 5: Ablation study on the influence of the masked conv, block masking, kernel size in stages 1
and 2 of ConvMAE on ImageNet-1K finetuning accuracy.

P-Epochs Method FT (%) LIN (%) AP box AP mask mIoU

ConvMAE-Base 84.1 N/A 50.2 44.8 48.1
200
w/ multi-scale decoder 84.4 N/A 50.8 45.4 48.5
ConvMAE-Base 84.6 69.4 52.5 46.5 50.7
1600
w/ multi-scale decoder 85.0 70.9 53.2 47.1 51.7
Table 6: For a base ConvMAE pretrained for 200 epochs and 1600 epochs, we ablate the multi-scale
decoder on ImageNet finetuning and object detection on COCO.

Convergence speed. We compare the convergence of ConvMAE and MAE in terms of ImageNet-1K
finetuning, linear probing accuracy and COCO AP box in Figure 4. For fair comparison, ConvMAE
and MAE are both pretrained for 1600 epochs. ConvMAE not only attains strong final results but

8
also significantly increases convergence speed on various tasks. Specifically, ConvMAE can surpass
the final performance of MAE at 58 epochs on ImageNet-1K finetuning. On COCO object detection,
ConvMAE surpasses MAE at 16 epochs, indicating 6.6× faster convergence speed.
4 Related Work
Vision Transformer. Vision Transformer(ViT) [18, 5] achieved state-of-the-art results on various
vision tasks. To increase the convergence speed and improve accuracy, well-explored locality
inductive bias have been reintroduced into vision transformer [66, 22, 62, 41, 27, 61, 51, 19, 56, 26],
among which, hybrid architecture of convolution and transformer design [49, 57, 12, 21, 34] can
achieve state-of-the-art performance of a wide range of tasks. Our ConvMAE is highly motivated
by the hybrid architecture design [21, 34, 12, 57] in vision backbones. Instead of designing new
architectures, ConvMAE aim to unleash the powerful representation induced by hybrid architectures
through MAE-style pretraining with several insightful modifications.
Self-supervised Representation Learning. Contrastive learning [8, 29, 6, 10] learn invariances
by comparing augmented views of un-labeled images. Recently Mask-Autoencoding motivated by
BERT [15] raised to be a promising methodology. Mask-Autoencoding can learn strong representation
through masked patch reconstruction with simple data augmentation. BEiT [2] introduced Mask-
Autoencoding into Vision Community. MAE [28] introduced an asymmetric encoder and decoder
architecture where masked tokens is skipped in computation-heavy encoder and only pass all tokens
through a light-weight decoder. iBoT [65] and Data2Vec [1], PeCo [17] and MaskFeat [55] explore
different reconstruction targets. Different from previous improvements of Mask-autoencoding,
ConvMAE introduce hierarchical representations architectures into MAE.
5 Conclusion
We propose a simple self-supervised learning framework named as ConvMAE which demonstrate
the hybrid local-global blocks [21, 34, 26, 19, 57, 49] can boost the performance of MAE [28] to
generate discriminative multi-scale features [38, 53, 42]. The computational efficiency and low
pretraing-fineuning gap of original MAE can be well maintained under our ConvMAE. ConvMAE
exhibits significantly improved performances on various vision tasks and can be easily implemented.
We will study combining improved reconstruction targets with ConvMAE in the future.
Negative societal impact: We do not foresee nagative social impact from the proposed work.

References
[1] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec:
A general framework for self-supervised learning in speech, vision and language. arXiv preprint
arXiv:2202.03555, 2022. 1, 5, 9
[2] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint
arXiv:2106.08254, 2021. 1, 2, 3, 4, 5, 6, 9
[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video
understanding. arXiv preprint arXiv:2102.05095, 2(3):4, 2021. 5
[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind
Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.
Advances in neural information processing systems, 33:1877–1901, 2020. 1
[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey
Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision,
pages 213–229. Springer, 2020. 9
[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand
Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF
International Conference on Computer Vision, pages 9650–9660, 2021. 1, 6, 9
[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset.
In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299–6308,
2017. 5
[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for
contrastive learning of visual representations. In International conference on machine learning, pages
1597–1607. PMLR, 2020. 9
[9] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping
Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning.
arXiv preprint arXiv:2202.03026, 2022. 6
[10] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision
transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages

9
 9640–9649, 2021. 1, 6, 9
[11] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, Xiaolin Wei, Huaxia Xia, and Chunhua Shen.
Conditional positional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 3
[12] Zihang Dai, Hanxiao Liu, Quoc Le, and Mingxing Tan. Coatnet: Marrying convolution and attention for
all data sizes. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 3, 9
[13] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE
computer society conference on computer vision and pattern recognition (CVPR’05), volume 1, pages
886–893. Ieee, 2005. 5
[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical
image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255.
Ieee, 2009. 1, 5
[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirec-
tional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1, 9
[16] Xiaohan Ding, Xiangyu Zhang, Yizhuang Zhou, Jungong Han, Guiguang Ding, and Jian Sun. Scaling up
your kernels to 31x31: Revisiting large kernel design in cnns. arXiv preprint arXiv:2203.06717, 2022. 8
[17] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang
Wen, and Nenghai Yu. Peco: Perceptual codebook for bert pre-training of vision transformers. arXiv
preprint arXiv:2111.12710, 2021. 6, 9
[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas
Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth
16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 9
[19] Stéphane d’Ascoli, Hugo Touvron, Matthew L Leavitt, Ari S Morcos, Giulio Biroli, and Levent Sagun.
Convit: Improving vision transformers with soft convolutional inductive biases. In International Conference
on Machine Learning, pages 2286–2296. PMLR, 2021. 1, 9
[20] Yuxin Fang, Shusheng Yang, Shijie Wang, Yixiao Ge, Ying Shan, and Xinggang Wang. Unleashing vanilla
vision transformer with masked image modeling for object detection. arXiv preprint arXiv:2204.02964,
2022. 2, 6
[21] Peng Gao, Jiasen Lu, Hongsheng Li, Roozbeh Mottaghi, and Aniruddha Kembhavi. Container: Context
aggregation network. arXiv preprint arXiv:2106.01401, 2021. 1, 2, 3, 9
[22] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr
with spatially modulated co-attention. In Proceedings of the IEEE/CVF International Conference on
Computer Vision, pages 3621–3630, 2021. 9
[23] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal,
Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The" something
something" video database for learning and evaluating visual common sense. In Proceedings of the IEEE
international conference on computer vision, pages 5842–5850, 2017. 7
[24] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with
submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision
and pattern recognition, pages 9224–9232, 2018. 2
[25] Benjamin Graham and Laurens van der Maaten. Submanifold sparse convolutional networks. arXiv
preprint arXiv:1706.01307, 2017. 2, 4
[26] Jianyuan Guo, Kai Han, Han Wu, Chang Xu, Yehui Tang, Chunjing Xu, and Yunhe Wang. Cmt: Convolu-
tional neural networks meet vision transformers. arXiv preprint arXiv:2107.06263, 2021. 3, 9
[27] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer.
Advances in Neural Information Processing Systems, 34, 2021. 9
[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders
are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 1, 2, 3, 4, 5, 6, 9, 12
[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised
visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern
recognition, pages 9729–9738, 2020. 5, 6, 9
[30] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE
international conference on computer vision, pages 2961–2969, 2017. 1, 2, 4, 5, 6
[31] Ajay Jain, Pieter Abbeel, and Deepak Pathak. Locally masked convolution for autoregressive models. In
Conference on Uncertainty in Artificial Intelligence, pages 1358–1367. PMLR, 2020. 2
[32] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan,
Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv
preprint arXiv:1705.06950, 2017. 7
[33] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional
neural networks. Advances in neural information processing systems, 25, 2012. 1
[34] Kunchang Li, Yali Wang, Junhao Zhang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu
Qiao. Uniformer: Unifying convolution and self-attention for visual recognition. arXiv preprint
arXiv:2201.09450, 2022. 1, 2, 3, 5, 9
[35] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones
for object detection. arXiv preprint arXiv:2203.16527, 2022. 6
[36] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph
Feichtenhofer. Improved multiscale vision transformers for classification and detection. arXiv preprint
arXiv:2112.01526, 2021. 2, 5, 6

10
[37] Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaiming He, and Ross Girshick. Benchmarking
detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429, 2021. 4, 6
[38] Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature
pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and
pattern recognition, pages 2117–2125, 2017. 4, 9
[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár,
and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on
computer vision, pages 740–755. Springer, 2014. 5
[40] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image
inpainting for irregular holes using partial convolutions. In Proceedings of the European conference on
computer vision (ECCV), pages 85–100, 2018. 2
[41] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr:
Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 9
[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin
transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF
International Conference on Computer Vision, pages 10012–10022, 2021. 1, 2, 3, 4, 5, 6, 9
[43] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A
convnet for the 2020s. arXiv preprint arXiv:2201.03545, 2022. 8
[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on
Learning Representations, 2018. 6
[45] Chao Peng, Xiangyu Zhang, Gang Yu, Guiming Luo, and Jian Sun. Large kernel matters–improve semantic
segmentation by global convolutional network. In Proceedings of the IEEE conference on computer vision
and pattern recognition, pages 4353–4361, 2017. 8
[46] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language
models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1
[47] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and
Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning,
pages 8821–8831. PMLR, 2021. 6
[48] Mengye Ren, Andrei Pokrovsky, Bin Yang, and Raquel Urtasun. Sbnet: Sparse blocks network for fast
inference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages
8711–8720, 2018. 2, 4
[49] Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani.
Bottleneck transformers for visual recognition. In Proceedings of the IEEE/CVF conference on computer
vision and pattern recognition, pages 16519–16529, 2021. 1, 3, 9
[50] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient
learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022. 5, 7
[51] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou.
Training data-efficient image transformers & distillation through attention. In International Conference on
Machine Learning, pages 10347–10357. PMLR, 2021. 6, 9
[52] Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby, Jakob Verbeek, and Hervé Jégou. Three things
everyone should know about vision transformers. arXiv preprint arXiv:2203.09795, 2022. 2
[53] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and
Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions.
In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568–578, 2021. 1, 9
[54] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In
Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794–7803, 2018.
5
[55] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked
feature prediction for self-supervised visual pre-training. arXiv preprint arXiv:2112.09133, 2021. 1, 5, 9
[56] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt:
Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF International Conference
on Computer Vision, pages 22–31, 2021. 9
[57] Tete Xiao, Piotr Dollar, Mannat Singh, Eric Mintun, Trevor Darrell, and Ross Girshick. Early convolutions
help transformers see better. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 3, 9
[58] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene
understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 418–434,
2018. 1, 2, 4, 6
[59] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu.
Simmim: A simple framework for masked image modeling. arXiv preprint arXiv:2111.09886, 2021. 2, 5
[60] Zhenda Xie, Zheng Zhang, Xizhou Zhu, Gao Huang, and Stephen Lin. Spatially adaptive inference with
stochastic feature sampling and interpolation. In European conference on computer vision, pages 531–548.
Springer, 2020. 2
[61] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng,
and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch on imagenet. In
Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 558–567, 2021. 9
[62] Gongjie Zhang, Zhipeng Luo, Yingchen Yu, Kaiwen Cui, and Shijian Lu. Accelerating detr convergence
via semantic-aligned matching. arXiv preprint arXiv:2203.06883, 2022. 9

11
[63] Bolei Zhou, Alex Andonian, Aude Oliva, and Antonio Torralba. Temporal relational reasoning in videos.
In Proceedings of the European conference on computer vision (ECCV), pages 803–818, 2018. 5
[64] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba.
Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision,
127(3):302–321, 2019. 6
[65] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image
bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 9
[66] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable
transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 9

6 Appendix
Architecture Details of ConvMAE Encoder. The details of our hybrid convolution-transformer
encoder is explained below. Given an input image I ∈ R3×H×W , stage 1 of ConvMAE encoder
H W
generates a high-resolution token embeddings E1 ∈ RC1 × 4 × 4 using non-overlapping 4 × 4 strided
convolution firstly. Then E1 is feed into stacked convolutional blocks which is repeated L1 times,
where L1 stands for the number of layers in stage 1. Similar as stage 1, stage 2 further downsamples
H W
feature map into token embeddings E2 ∈ RC2 × 8 × 8 using non-overlapping 2×2 strided convolution.
E2 is processed by L2 layers of convolutional blocks again. After local information fusion utilized in
stage 1 and stage 2, stage 3 perform global feature fusion using transformer block. E2 is projected into
H W
tokens embeddings E3 ∈ R( 16 × 16 )×C3 using non-overlapping 2 × 2 strided convolution. E3 mixing
with Intermediate Positional Embedding (IPL) is feed into a pure transformer block with L3 layers.
We denote the number of attention heads in stage 3 as Ha . The mlp-ratios in FFN for different stages
is denoted as P1 , P2 and P3 in respectively. Stage 1 and stage 2 is designed to capture fine-grained
details on high resolution feature map. Stage 3 can perform dynamically global reasoning efficiently
on a rather low-resolution feature map. At the same time, stage 3 can enlarge the filed-of-view
(FOV) of backbone which benefits a wide range of downstream tasks. The encoder of ConvMAE
can seamlessly inherits the merits of convolution and transformer block. The architecture details for
small, base and large model is listed in Table 7. ConvMAE small, base, large and huge share similar
parameter scale with the encoder of MAE-small, MAE-base, MAE-large and MAE-huge.

Model [C1 , C2 , C3 ] [L1 , L2 , L3 ] [E1 , E2 , E3 ] [P1 , P2 , P3 ] Ha #Params (M)

ConvMAE-S [128, 256, 384] [2, 2, 11] [56, 28, 14] [4, 4, 4] 6 22
ConvMAE-B [256, 384, 768] [2, 2, 11] [56, 28,14] [4, 4, 4] 12 84
ConvMAE-B* [256, 384, 768] [2, 2, 11] [56, 28,14] [8, 8, 4] 12 88
ConvMAE-L [384, 768, 1024] [2, 2, 23] [56, 28, 14] [8, 8, 4] 16 322
ConvMAE-H [768, 1024, 1280] [2, 2, 31] [56, 28, 14] [8, 8, 4] 16 666

Table 7: Architecture details of ConvMAE small, base, large and huge. ConvMAE-B* repre-
sents multi-scale encoder with large mlp-ratios in stage 1 and stage 2. [C1 , C2 , C3 ], [L1 , L2 , L3 ],
[E1 , E2 , E3 ] and [P1 , P2 , P3 ] represents channel dimension, number of layer, spatial resolution and
mlp-ratios for each stage 1, stage 2 and stage 3. Ha stands for the number of attention heads in stage
3.

Model Scaling up and down. We design ConvMAE of different parameters scales to match those of
MAE-small, MAE-base, MAE-large and MAE-huge. Detailed network architectures are in appendix.
The finetuning performances are shown in Table 8. Compared with the original MAE [28] of different
scales, our ConvMAE of different scales consistently outperform its MAE counterparts on Imagenet
finetuning. This suggests that ConvMAE can be an efficient learner for different paramter scales.

Model size
Method P-Epochs
Small Base Base* Large Huge
MAE 1600 79.5 83.6 N/A 85.9 86.9
ConvMAE 800 82.6 84.6 84.9 86.2 N/A

Table 8: Ablation study of model scales.

12