Sequence to Sequence Video to Text

Subhashini Venugopalan1

Marcus Rohrbach2,4
Trevor Darrell2

arXiv:1505.00487v3 [cs.CV] 19 Oct 2015

University of Texas at Austin

University of Massachusetts, Lowell

Jeff Donahue2
Kate Saenko3

Raymond Mooney1

University of California, Berkeley

International Computer Science Institute, Berkeley

Abstract
Raw Frames

Real-world videos often have complex dynamics; and

methods for generating open-domain video descriptions
should be sensitive to temporal structure and allow both input (sequence of frames) and output (sequence of words) of
variable length. To approach this problem, we propose a
novel end-to-end sequence-to-sequence model to generate
captions for videos. For this we exploit recurrent neural networks, specifically LSTMs, which have demonstrated stateof-the-art performance in image caption generation. Our
LSTM model is trained on video-sentence pairs and learns
to associate a sequence of video frames to a sequence of
words in order to generate a description of the event in the
video clip. Our model naturally is able to learn the temporal structure of the sequence of frames as well as the sequence model of the generated sentences, i.e. a language
model. We evaluate several variants of our model that exploit different visual features on a standard set of YouTube
videos and two movie description datasets (M-VAD and
MPII-MD).

CNN - Object
pretrained

CNN
Outputs

LSTMs

Our LSTM network is connected to a

CNN for RGB frames or a
CNN for optical flow images.

A
man
is

Flow images

cutting
a
CNN - Action
pretrained

bottle
<eos>

Figure 1. Our S2VT approach performs video description using

a sequence to sequence model. It incorporates a stacked LSTM
which first reads the sequence of frames and then generates a sequence of words. The input visual sequence to the model is comprised of RGB and/or optical flow CNN outputs.

sequentially and outputs words sequentially.

1. Introduction

The problem of generating descriptions in open domain

videos is difficult not just due to the diverse set of objects,
scenes, actions, and their attributes, but also because it is
hard to determine the salient content and describe the event
appropriately in context. To learn what is worth describing, our model learns from video clips and paired sentences
that describe the depicted events in natural language. We
use Long Short Term Memory (LSTM) networks [12], a
type of recurrent neural network (RNN) that has achieved
great success on similar sequence-to-sequence tasks such as
speech recognition [10] and machine translation [34]. Due
to the inherent sequential nature of videos and language,
LSTMs are well-suited for generating descriptions of events
in videos.

Describing visual content with natural language text has

recently received increased interest, especially describing
images with a single sentence [8, 5, 16, 18, 20, 23, 29, 40].
Video description has so far seen less attention despite its
important applications in human-robot interaction, video indexing, and describing movies for the blind. While image
description handles a variable length output sequence of
words, video description also has to handle a variable length
input sequence of frames. Related approaches to video
description have resolved variable length input by holistic
video representations [29, 28, 11], pooling over frames [39],
or sub-sampling on a fixed number of input frames [43]. In
contrast, in this work we propose a sequence to sequence
model which is trained end-to-end and is able to learn arbitrary temporal structure in the input sequence. Our model
is sequence to sequence in a sense that it reads in frames

The main contribution of this work is to propose a novel

model, S2VT, which learns to directly map a sequence of
1

frames to a sequence of words. Figure 1 depicts our model.

A stacked LSTM first encodes the frames one by one, taking as input the output of a Convolutional Neural Network
(CNN) applied to each input frames intensity values. Once
all frames are read, the model generates a sentence word
by word. The encoding and decoding of the frame and
word representations are learned jointly from a parallel corpus. To model the temporal aspects of activities typically
shown in videos, we also compute the optical flow [2] between pairs of consecutive frames. The flow images are also
passed through a CNN and provided as input to the LSTM.
Flow CNN models have been shown to be beneficial for activity recognition [31, 8].
To our knowledge, this is the first approach to video description that uses a general sequence to sequence model.
This allows our model to (a) handle a variable number of
input frames, (b) learn and use the temporal structure of the
video and (c) learn a language model to generate natural,
grammatical sentences. Our model is learned jointly and
end-to-end, incorporating both intensity and optical flow
inputs, and does not require an explicit attention model.
We demonstrate that S2VT achieves state-of-the-art performance on three diverse datasets, a standard YouTube corpus (MSVD) [3] and the M-VAD [37] and MPII Movie
Description [28] datasets. Our implementation (based on
the Caffe [15] deep learning framework) is available on
github. https://github.com/vsubhashini/caffe/
tree/recurrent/examples/s2vt.

2. Related Work
Early work on video captioning considered tagging
videos with metadata [1] and clustering captions and videos
[14, 25, 42] for retrieval tasks. Several previous methods
for generating sentence descriptions [11, 19, 36] used a two
stage pipeline that first identifies the semantic content (subject, verb, object) and then generates a sentence based on a
template. This typically involved training individual classifiers to identify candidate objects, actions and scenes. They
then use a probabilistic graphical model to combine the visual confidences with a language model in order to estimate
the most likely content (subject, verb, object, scene) in the
video, which is then used to generate a sentence. While this
simplified the problem by detaching content generation and
surface realization, it requires selecting a set of relevant objects and actions to recognize. Moreover, a template-based
approach to sentence generation is insufficient to model the
richness of language used in human descriptions e.g.,
which attributes to use and how to combine them effectively to generate a good description. In contrast, our approach avoids the separation of content identification and
sentence generation by learning to directly map videos to
full human-provided sentences, learning a language model
simultaneously conditioned on visual features.

Our models take inspiration from the image caption generation models in [8, 40]. Their first step is to generate a
fixed length vector representation of an image by extracting features from a CNN. The next step learns to decode
this vector into a sequence of words composing the description of the image. While any RNN can be used in principle
to decode the sequence, the resulting long-term dependencies can lead to inferior performance. To mitigate this issue,
LSTM models have been exploited as sequence decoders, as
they are more suited to learning long-range dependencies.
In addition, since we are using variable-length video as input, we use LSTMs as sequence to sequence transducers,
following the language translation models of [34].
In [39], LSTMs are used to generate video descriptions
by pooling the representations of individual frames. Their
technique extracts CNN features for frames in the video and
then mean-pools the results to get a single feature vector
representing the entire video. They then use an LSTM as
a sequence decoder to generate a description based on this
vector. A major shortcoming of this approach is that this
representation completely ignores the ordering of the video
frames and fails to exploit any temporal information. The
approach in [8] also generates video descriptions using an
LSTM; however, they employ a version of the two-step approach that uses CRFs to obtain semantic tuples of activity,
object, tool, and locatation and then use an LSTM to translate this tuple into a sentence. Moreover, the model in [8] is
applied to the limited domain of cooking videos while ours
is aimed at generating descriptions for videos in the wild.
Contemporaneous with our work, the approach in [43]
also addresses the limitations of [39] in two ways. First,
they employ a 3-D convnet model that incorporates spatiotemporal motion features. To obtain the features, they assume videos are of fixed volume (width, height, time). They
extract dense trajectory features (HoG, HoF, MBH) [41]
over non-overlapping cuboids and concatenate these to form
the input. The 3-D convnet is pre-trained on video datasets
for action recognition. Second, they include an attention
mechanism that learns to weight the frame features nonuniformly conditioned on the previous word input(s) rather
than uniformly weighting features from all frames as in
[39]. The 3-D convnet alone provides limited performance
improvement, but in conjunction with the attention model it
notably improves performance. We propose a simpler approach to using temporal information by using an LSTM
to encode the sequence of video frames into a distributed
vector representation that is sufficient to generate a sentential description. Therefore, our direct sequence to sequence
model does not require an explicit attention mechanism.
Another recent project [33] uses LSTMs to predict the
future frame sequence from an encoding of the previous
frames. Their model is more similar to the language translation model in [34], which uses one LSTM to encode the

input text into a fixed representation, and another LSTM to

decode it into a different language. In contrast, we employ a
single LSTM that learns both encoding and decoding based
on the inputs it is provided. This allows the LSTM to share
weights between encoding and decoding.
Other related work includes [24, 8], which uses LSTMs
for activity classification, predicting an activity class for the
representation of each image/flow frame. In contrast, our
model generates captions after encoding the complete sequence of optical flow images. Specifically, our final model
is an ensemble of the sequence to sequence models trained
on raw images and optical flow images.

3. Approach
We propose a sequence to sequence model for video description, where the input is the sequence of video frames
(x1 , . . . , xn ), and the output is the sequence of words
(y1 , . . . , ym ). Naturally, both the input and output are of
variable, potentially different, lengths. In our case, there
are typically many more frames than words.
In our model, we estimate the conditional probability of
an output sequence (y1 , . . . , ym ) given an input sequence
(x1 , . . . , xn ) i.e.
p(y1 , . . . , ym |x1 , . . . , xn )
(1)
This problem is analogous to machine translation between
natural languages, where a sequence of words in the input
language is translated to a sequence of words in the output
language. Recently, [6, 34] have shown how to effectively
attack this sequence to sequence problem with an LSTM
Recurrent Neural Network (RNN). We extend this paradigm
to inputs comprised of sequences of video frames, significantly simplifying prior RNN-based methods for video description. In the following, we describe our model and architecture in detail, as well as our input and output representation for video and sentences.

3.1. LSTMs for sequence modeling

The main idea to handle variable-length input and output
is to first encode the input sequence of frames, one at a time,
representing the video using a latent vector representation,
and then decode from that representation to a sentence, one
word at a time.
Let us first recall the Long Short Term Memory RNN
(LSTM), originally proposed in [12]. Relying on the LSTM
unit proposed in [44], for an input xt at time step t, the
LSTM computes a hidden/control state ht and a memory
cell state ct which is an encoding of everything the cell has
observed until time t:
it = (Wxi xt + Whi ht1 + bi )
ft = (Wxf xt + Whf ht1 + bf )
ot = (Wxo xt + Who ht1 + bo )
(2)
gt = (Wxg xt + Whg ht1 + bg )
ct = ft  ct1 + it  gt
ht = ot  (ct )

where is the sigmoidal non-linearity, is the hyperbolic

tangent non-linearity,  represents the element-wise product with the gate value, and the weight matrices denoted by
Wij and biases bj are the trained parameters.
Thus, in the encoding phase, given an input sequence
X (x1 , . . . , xn ), the LSTM computes a sequence of hidden
states (h1 , . . . , hn ). During decoding it defines a distribution over the output sequence Y (y1 , . . . , ym ) given the input sequence X as p(Y |X) is
m
Y
p(y1 , . . . , ym |x1 , . . . , xn ) =
p(yt |hn+t1 , yt1 ) (3)
t=1

where the distribution of p(yt |hn+t ) is given by a softmax

over all of the words in the vocabulary (see Equation 5).
Note that hn+t is obtained from hn+t1 , yt1 based on the
recursion in Equation 2.

3.2. Sequence to sequence video to text

Our approach, S2VT, is depicted in Figure 2. While
[6, 34] first encode the input sequence to a fixed length vector using one LSTM and then use another LSTM to map the
vector to a sequence of outputs, we rely on a single LSTM
for both the encoding and decoding stage. This allows parameter sharing between the encoding and decoding stage.
Our model uses a stack of two LSTMs with 1000 hidden units each. Figure 2 shows the LSTM stack unrolled
over time. When two LSTMs are stacked together, as in our
case, the hidden representation (ht ) from the first LSTM
layer (colored red) is provided as the input (xt ) to the second LSTM (colored green). The top LSTM layer in our architecture is used to model the visual frame sequence, and
the next layer is used to model the output word sequence.
Training and Inference In the first several time steps,
the top LSTM layer (colored red in Figure 2) receives a sequence of frames and encodes them while the second LSTM
layer receives the hidden representation (ht ) and concatenates it with null padded input words (zeros), which it then
encodes. There is no loss during this stage when the LSTMs
are encoding. After all the frames in the video clip are exhausted, the second LSTM layer is fed the beginning-ofsentence (<BOS>) tag, which prompts it to start decoding
its current hidden representation into a sequence of words.
While training in the decoding stage, the model maximizes
for the log-likelihood of the predicted output sentence given
the hidden representation of the visual frame sequence, and
the previous words it has seen. From Equation 3 for a model
with parameters and output sequence Y = (y1 , . . . , ym ),
this is formulated as:
m
X

= argmax
log p(yt |hn+t1 , yt1 ; )
(4)

t=1

This log-likelihood is optimized over the entire training

dataset using stochastic gradient descent. The loss is computed only when the LSTM is learning to decode. Since this

LSTM

<pad>

LSTM

<pad>
LSTM

LSTM

<pad>
LSTM

LSTM

<pad>
LSTM

<pad>

<pad>

<pad>

<pad>

<pad>

LSTM

LSTM

LSTM

LSTM

LSTM

LSTM

LSTM

LSTM

LSTM

LSTM

man

is

talking

<EOS>

<BOS>
LSTM

Encoding stage

Decoding stage

time

Figure 2. We propose a stack of two LSTMs that learn a representation of a sequence of frames in order to decode it into a sentence that
describes the event in the video. The top LSTM layer (colored red) models visual feature inputs. The second LSTM layer (colored green)
models language given the text input and the hidden representation of the video sequence. We use <BOS> to indicate begin-of-sentence
and <EOS> for the end-of-sentence tag. Zeros are used as a <pad> when there is no input at the time step.

loss is propagated back in time, the LSTM learns to generate an appropriate hidden state representation (hn ) of the
input sequence. The output (zt ) of the second LSTM layer
is used to obtain the emitted word (y). We apply a softmax
function to get the probability distribution over the words y 0
in the vocabulary V :
p(y|zt ) = P

exp(Wy zt )
exp(Wy0 zt )

(5)

y 0 V

We note that, during the decoding phase, the visual frame

representation for the first LSTM layer is simply a vector
of zeros that acts as padding input. We require an explicit
end-of-sentence tag (<EOS>) to terminate each sentence
since this enables the model to define a distribution over
sequences of varying lengths. At test time, during each decoding step we choose the word yt with the maximum probability after the softmax (from Equation 5) until it emits the
<EOS> token.

3.3. Video and text representation

RGB frames. Similar to previous LSTM-based image captioning efforts [8, 40] and video-to-text approaches [39, 43],
we apply a convolutional neural network (CNN) to input
images and provide the output of the top layer as input to
the LSTM unit. In this work, we report results using the output of the fc7 layer (after applying the ReLU non-linearity)
on the Caffe Reference Net (a variant of AlexNet) and also
the 16-layer VGG model [32]. We use CNNs that are pretrained on the 1.2M image ILSVRC-2012 object classification subset of the ImageNet dataset [30] and made available
publicly via the Caffe ModelZoo.1 Each input video frame
is scaled to 256x256, and is cropped to a random 227x227
1 https://github.com/BVLC/caffe/wiki/Model-Zoo

region. It is then processed by the CNN. We remove the

original last fully-connected classification layer and learn a
new linear embedding of the features to a 500 dimensional
space. The lower dimension features form the input (xt )
to the first LSTM layer. The weights of the embedding are
learned jointly with the LSTM layers during training.
Optical Flow. In addition to CNN outputs from raw image (RGB) frames, we also incorporate optical flow measures as input sequences to our architecture. Others [24, 8]
have shown that incorporating optical flow information to
LSTMs improves activity classification. As many of our
descriptions are activity centered, we explore this option
for video description as well. We follow the approach in
[8, 9] and first extract classical variational optical flow features [2]. We then create flow images (as seen in Figure 1)
in a manner similar to [9], by centering x and y flow values
around 128 and multiplying by a scalar such that flow values
fall between 0 and 255. We also calculate the flow magnitude and add it as a third channel to the flow image. We
then use a CNN [9] initialized with weights trained on the
UCF101 video dataset to classify optical flow images into
101 activity classes. The fc6 layer activations of the CNN
are embedded in a lower 500 dimensional space which is
then given as input to the LSTM. The rest of the LSTM architecture remains unchanged for flow inputs.
In our combined model, we use a shallow fusion technique to integrate flow and RGB features. At each time
step of the decoding phase, the model proposes a set of candidate words. We then rescore these hypotheses with the
weighted sum of the scores by the flow and RGB networks,
where we only need to recompute the score of each new
word p(yt = y 0 ) as:
prgb (yt = y 0 ) + (1 ) pf low (yt = y 0 )
the hyper-parameter is tuned on the validation set.

#-sentences
#-tokens
vocab
#-videos
avg. length
#-sents per video

MSVD

MPII-MD

MVAD

80,827
567,874
12,594
1,970
10.2s
41

68,375
679,157
21,700
68,337
3.9s
1

56,634
568,408
18,092
46,009
6.2s
1-2

Table 1. Corpus Statistics. The the number of tokens in all datasets

are comparable, however MSVD has multiple descriptions for
each video while the movie corpora (MPII-MD, MVAD) have a
large number of clips with a single description each. Thus, the
number of video-sentence pairs in all 3 datasets are comparable.

Text input. The target output sequence of words are

represented using one-hot vector encoding (1-of-N coding,
where N is the size of the vocabulary). Similar to the treatment of frame features, we embed words to a lower 500
dimensional space by applying a linear transformation to
the input data and learning its parameters via back propagation. The embedded word vector concatenated with the
output (ht ) of the first LSTM layer forms the input to the
second LSTM layer (marked green in Figure 2). When considering the output of the LSTM we apply a softmax over
the complete vocabulary as in Equation 5.

verting all text to lower case, tokenizing the sentences and

removing punctuation. We use the data splits provided by
[39]. Additionally, in each video, we sample every tenth
frame as done by [39].
4.1.2

MPII Movie Description Dataset (MPII-MD)

MPII-MD [28] contains around 68,000 video clips extracted

from 94 Hollywood movies. Each clip is accompanied with
a single sentence description which is sourced from movie
scripts and audio description (AD) data. AD or Descriptive Video Service (DVS) is an additional audio track that
is added to the movies to describe explicit visual elements
in a movie for the visually impaired. Although the movie
snippets are manually aligned to the descriptions, the data is
very challenging due to the high diversity of visual and textual content, and the fact that most snippets have only a single reference sentence. We use the training/validation/test
split provided by the authors and extract every fifth frame
(videos are shorter than MSVD, averaging 94 frames).
4.1.3

Montreal Video Annotation Dataset (M-VAD)

The M-VAD movie description corpus [37] is another recent

collection of about 49,000 short video clips from 92 movies.
It is similar to MPII-MD, but contains only AD data with
automatic alignment. We use the same setup as for MPIIMD.

4. Experimental Setup

4.2. Evaluation Metrics

This secction describes the evaluation of our approach.

We first describe the datasets used, then the evaluation protocol, and then the details of our models.

Quantitative evaluation of the models are performed using the METEOR [7] metric which was originally proposed to evaluate machine translation results. The METEOR score is computed based on the alignment between
a given hypothesis sentence and a set of candidate reference sentences. METEOR compares exact token matches,
stemmed tokens, paraphrase matches, as well as semantically similar matches using WordNet synonyms. This semantic aspect of METEOR distinguishes it from others such
as BLEU [26], ROUGE-L [21], or CIDEr [38]. The authors of CIDEr [38] evaluated these four measures for image description. They showed that METEOR is always better than BLEU and ROUGE and outperforms CIDEr when
the number of references are small (CIDEr is comparable to
METEOR when the number of references are large). Since
MPII-MD and M-VAD have only a single reference, we decided to use METEOR in all our evaluations. We employ
METEOR version 1.5 2 using the code3 released with the
Microsoft COCO Evaluation Server [4].

4.1. Video description datasets

We report results on three video description corpora,
namely the Microsoft Video Description corpus (MSVD)
[3], the MPII Movie Description Corpus (MPII-MD) [28],
and the Montreal Video Annotation Dataset (M-VAD)
[37]. Together they form the largest parallel corpora
with open domain video and natural language descriptions.
While MSVD is based on web clips with short humanannotated sentences, MPII-MD and M-VAD contain Hollywood movie snippets with descriptions sourced from script
data and audio description. Statistics of each corpus are
presented in Table 1.
4.1.1

Microsoft Video Description Corpus (MSVD)

The Microsoft Video description corpus [3], is a collection

of Youtube clips collected on Mechanical Turk by requesting workers to pick short clips depicting a single activity.
The videos were then used to elicit single sentence descriptions from annotators. The original corpus has multi-lingual
descriptions, in this work we use only the English descriptions. We do minimal pre-processing on the text by con-

4.3. Experimental details of our models

All our models take as input either the raw RGB frames
directly feeding into the CNN, or pre-processed optical flow
2 http://www.cs.cmu.edu/

alavie/METEOR

3 https://github.com/tylin/coco-caption

images (described in Section 3.3). In all of our models, we

unroll the LSTM to a fixed 80 time steps during training.
We found this to be a good trade-off between memory consumption and the ability to provide many frames (videos) to
the LSTM. This setting allows us to fit multiple videos in a
single mini-batch (up to 8 for AlexNet and up to 3 for flow
models). We note that 94% of the YouTube training videos
satisfied this limit (with frames sampled at the rate of 1 in
10). For videos with fewer than 80 time steps (of words
and frames), we pad the remaining inputs with zeros. For
longer videos, we truncate the number of frames to ensure
that the sum of the number of frames and words is within
this limit. At test time, we do not constrain the length of the
video and our model views all sampled frames. We use the
pre-trained AlexNet and VGG CNNs. For VGG, we fix all
layers below fc7 to reduce memory consumption and allow
faster training.
We compare our sequence to sequence LSTM architecture with RGB image features extracted from both AlexNet,
and the 16-layer VGG network. In order to compare features from the VGG network with previous models, we include the performance of the mean-pooled model proposed
in [39] using the output of the fc7 layer from the 16 layer
VGG as a baseline (line 3, Table 2). All our sequence to sequence models are referenced in Table 2 under S2VT. Our
first variant, RGB (AlexNet) is the end-to-end model that
uses AlexNet on RGB frames. Flow (AlexNet) refers to the
model that is obtained by training on optical flow images.
RGB (VGG) refers to the model with the 16-layer VGG
model on RGB image frames. We also experiment with randomly re-ordered input frames (line 10) to verify that S2VT
learns temporal-sequence information. Our final model is
an ensemble of the RGB (VGG) and Flow (AlexNet) where
the prediction at each time step is a weighted average of the
prediction from the individual models.

4.4. Related approaches

We compare our sequence to sequence models against
the factor graph model (FGM) in [36], the mean-pooled
models in [39] and the Soft-Attention models of [43].
FGM proposed in [36] uses a two step approach to first
obtain confidences on subject, verb, object and scene
(S,V,O,P) elements and combines these with confidences
from a language model using a factor graph to infer the most
likely (S,V,O,P) tuple in the video. It then generates a sentence based on a template.
The Mean Pool model proposed in [39] pools AlexNet fc7
activations across all frames to create a fixed-length vector
representation of the video. It then uses an LSTM to then
decode the vector into a sequence of words. Further, the
model ia pre-trained on the Flickr30k [13] and MSCOCO
[22] image-caption datasets and fine-tuned on MSVD for
a significant improvement in performance. We compare

Model
FGM [36]
Mean pool
- AlexNet [39]
- VGG
- AlexNet COCO pre-trained [39]
- GoogleNet [43]
Temporal attention
- GoogleNet [43]
- GoogleNet + 3D-CNN [43]
S2VT (ours)
- Flow (AlexNet)
- RGB (AlexNet)
- RGB (VGG) random frame order
- RGB (VGG)
- RGB (VGG) + Flow (AlexNet)

METEOR
23.9

(1)

26.9
27.7
29.1
28.7

(2)
(3)
(4)
(5)

29.0
29.6

(6)
(7)

24.3
27.9
28.2
29.2
29.8

(8)
(9)
(10)
(11)
(12)

Table 2. MSVD dataset (METEOR in %, higher is better).

our models against their basic mean-pooled model and their

best model obtained from fine-tuning on Flickr30k and
COCO datasets. We also compare against the GoogleNet
[35] variant of the mean-pooled model reported in [43].
The Temporal-Attention model in [43] is a combination of
weighted attention over a fixed set of video frames with input features from GoogleNet and a 3D-convnet trained on
HoG, HoF and MBH features from an activity classification
model.

5. Results and Discussion

This section discussses the result of our evaluation
shown in Tables 2, 4, and 5.

5.1. MSVD dataset

Table 2 shows the results on the MSVD dataset. Rows
1 through 7 present related approaches and the rest are
variants of our S2VT approach. Our basic S2VT AlexNet
model on RGB video frames (line 9 in Table 2) achieves
27.9% METEOR and improves over the basic mean-pooled
model in [39] (line 2, 26.9%) as well as the VGG meanpooled model (line 3, 27.7%);suggesting that S2VT is a
more powerful approach. When the model is trained with
randomly-ordered frames (line 10 in Table 2), the score is
considerably lower, clearly demonstrating that the model
benefits from exploiting temporal structure.
Our S2VT model which uses flow images (line 8)
achieves only 24.3% METEOR but improves the performance of our VGG model from 29.2% (line 11) to 29.8%
(line 12), when combined. A reason for the low performance of the flow model could be that optical flow features even for the same activity can vary significantly with
context e.g. panda eating vs person eating. Also, the

Edit-Distance
MSVD
MPII-MD
MVAD

k=0

k <= 1

k <= 2

k <= 3

42.9
28.8
15.6

81.2
43.5
28.7

93.6
56.4
37.8

96.6
83.0
45.0

Table 3. Percentage of generated sentences which match a sentence of the training set with an edit (Levenshtein) distance of less
than 4. All values reported in percentage (%).

model only receives very weak signals with regard to the

kind of activities depicted in YouTube videos. Some commonly used verbs such as play are polysemous and can
refer to playing a musical instrument (playing a guitar) or
playing a sport (playing golf). However, integrating RGB
with Flow improves the quality of descriptions.
Our ensemble using both RGB and Flow performs
slightly better than the best model proposed in [43], temporal attention with GoogleNet + 3D-CNN (line 7). The
modest size of the improvement is likely due to the much
stronger 3D-CNN features (as the difference to GoogleNet
alone (line 6) suggests). Thus, the closest comparison between the Temporal Attention Model [43] and S2VT is arguably S2VT with VGG (line 12) vs. their GoogleNet-only
model (line 6).
Figure 3 shows descriptions generated by our model on
sample Youtube clips from MSVD. To compare the originality in generation, we compute the Levenshtein distance
of the predicted sentences with those in the training set.
From Table 3, for the MSVD corpus, 42.9% of the predictions are identical to some training sentence, and another
38.3% can be obtained by inserting, deleting or substituting
one word from some sentence in the training corpus. We
note that many of the descriptions generated are relevant.

5.2. Movie description datasets

For the more challenging MPII-MD and M-VAD
datasets we use our single best model, namely S2VT trained
on RGB frames and VGG. To avoid over-fitting on the
movie corpora we employ drop-out which has proved to be
beneficial on these datasets [27]. We found it was best to
use dropout at the inputs and outputs of both LSTM layers. Further, we used ADAM [17] for optimization with a
first momentum coefficient of 0.9 and a second momentum
coefficient of 0.999. For MPII-MD, reported in Table 4,
we improve over the SMT approach from [28] from 5.6%
to 7.1% METEOR and over Mean pooling [39] by 0.4%.
Our performance is similar to Visual-Labels [27], a contemporaneous LSTM-based approach which uses no temporal
encoding, but more diverse visual features, namely object
detectors, as well as activity and scene classifiers.
On M-VAD we achieve 6.7% METEOR which significantly outperforms the temporal attention model [43]

Approach
SMT (best variant) [28]
Visual-Labels [27]
Mean pool (VGG)
S2VT: RGB (VGG), ours

METEOR
5.6
7.0
6.7
7.1

Table 4. MPII-MD dataset (METEOR in %, higher is better).

Approach
Visual-Labels [27]
Temporal att. (GoogleNet+3D-CNN) [43] 4
Mean pool (VGG)
S2VT: RGB (VGG), ours

METEOR
6.3
4.3
6.1
6.7

Table 5. M-VAD dataset (METEOR in %, higher is better).

(4.3%)4 and Mean pooling (6.1%). On this dataset we also

outperform Visual-Labels [27] (6.3%).
We report results on the LSMDC challenge5 , which
combines M-VAD and MPII-MD. S2VT achieves 7.0%
METEOR on the public test set using the evaluation server.
In Figure 4 we present descriptions generated by our
model on some sample clips from the M-VAD dataset.
More example video clips, generated sentences, and data
are available on the authors webpages6 .

6. Conclusion
This paper proposed a novel approach to video description. In contrast to related work, we construct descriptions using a sequence to sequence model, where frames
are first read sequentially and then words are generated sequentially. This allows us to handle variable-length input
and output while simultaneously modeling temporal structure. Our model achieves state-of-the-art performance on
the MSVD dataset, and outperforms related work on two
large and challenging movie-description datasets. Despite
its conceptual simplicity, our model significantly benefits
from additional data, suggesting that it has a high model
capacity, and is able to learn complex temporal structure
in the input and output sequences for challenging moviedescription datasets.

Acknowledgments
We thank Lisa Anne Hendricks, Matthew Hausknecht,
Damian Mrowca for helpful discussions; and Anna
Rohrbach for help with both movie corpora; and the
4 We report results using the predictions provided by [43] but using the
orginal COCO Evaluation scripts. [43] report 5.7% METEOR for their
temporal attention + 3D-CNN model using a different tokenization.
5 LSMDC: sites.google.com/site/describingmovies
6 http://vsubhashini.github.io/s2vt.html

Correct descriptions.

Relevant but incorrect

descriptions.

Irrelevant descriptions.

(a)

(b)

(c)

Figure 3. Qualitative results on MSVD YouTube dataset from our S2VT model (RGB on VGG net). (a) Correct descriptions involving
different objects and actions for several videos. (b) Relevant but incorrect descriptions. (c) Descriptions that are irrelevant to the event in
the video.

(1)

(2)

Temporal Attention (GNet+3D-convatt):

(1) At night , SOMEONE and SOMEONE
step into the parking lot.
(2) Now the van drives away.
(3) They drive away.
(4) They drive off.
(5) They drive off.
(6) At the end of the street , SOMEONE
sits with his eyes closed.

(3)

(4)

(5)

(6a)

(6b)

S2VT (Ours): (1) Now, the van pulls out a window and a

DVS: (1) Now , at night , our view glides over a highway ,

tall brick facade of tall trees . a figure stands at a curb.

(2) Someone drives off the passenger car and drives off.

its lanes glittering from the lights of traffic below.

(3) They drive off the street.

(4) They drive off a suburban road and parks in a dirt
neighborhood.
(5) They drive off a suburban road and parks on a street.
(6) Someone sits in the doorway and stares at her
with a furrowed brow.

(2) Someone's suv cruises down a quiet road.

(3) Then turn into a parking lot .
(4) A neon palm tree glows on a sign that reads
oasis motel.
(5) Someone parks his suv in front of some rooms.
(6) He climbs out with his briefcase , sweeping his
cautious gaze around the area.

Figure 4. M-VAD Movie corpus: Representative frame from 6 contiguous clips from the movie Big Mommas: Like Father, Like Son.
From left: Temporal Attention (GoogleNet+3D-CNN) [43], S2VT (in blue) trained on the M-VAD dataset, and DVS: ground truth.

anonymous reviewers for insightful comments and suggestions. We acknowledge support from ONR ATL Grant
N00014-11-1-010, DARPA, AFRL, DoD MURI award
N000141110688, DEFT program (AFRL grant FA8750-132-0026), NSF awards IIS-1427425, IIS-1451244, and IIS1212798, and Berkeley Vision and Learning Center. Raymond and Kate acknowledge support from Google. Marcus
was supported by the FITweltweit-Program of the German
Academic Exchange Service (DAAD).

References
[1] H. Aradhye, G. Toderici, and J. Yagnik. Video2text: Learning to annotate video content. In ICDMW, 2009. 2
[2] T. Brox, A. Bruhn, N. Papenberg, and J. Weickert. High accuracy optical flow estimation based on a theory for warping.
In ECCV, pages 2536, 2004. 2, 4
[3] D. L. Chen and W. B. Dolan. Collecting highly parallel data
for paraphrase evaluation. In ACL, 2011. 2, 5
[4] X. Chen, H. Fang, T.-Y. Lin, R. Vedantam, S. Gupta, P. Dol-

[5]
[6]

[7]

[8]

[9]
[10]
[11]

[12]
[13]

[14]
[15]

[16]
[17]
[18]

[19]

[20]

[21]

[22]

[23]

lar, and C. L. Zitnick. Microsoft COCO captions: Data collection and evaluation server. arXiv:1504.00325, 2015. 5
X. Chen and C. L. Zitnick. Learning a recurrent visual representation for image caption generation. CVPR, 2015. 1
K. Cho, B. van Merrienboer, D. Bahdanau, and Y. Bengio.
On the properties of neural machine translation: Encoderdecoder approaches. arXiv:1409.1259, 2014. 3
M. Denkowski and A. Lavie. Meteor universal: Language
specific translation evaluation for any target language. In
EACL, 2014. 5
J. Donahue, L. A. Hendricks, S. Guadarrama, M. Rohrbach,
S. Venugopalan, K. Saenko, and T. Darrell. Long-term recurrent convolutional networks for visual recognition and description. In CVPR, 2015. 1, 2, 3, 4
G. Gkioxari and J. Malik. Finding action tubes. 2014. 4
A. Graves and N. Jaitly. Towards end-to-end speech recognition with recurrent neural networks. In ICML, 2014. 1
S. Guadarrama, N. Krishnamoorthy, G. Malkarnenkar,
S. Venugopalan, R. Mooney, T. Darrell, and K. Saenko.
Youtube2text: Recognizing and describing arbitrary activities using semantic hierarchies and zero-shoot recognition.
In ICCV, 2013. 1, 2
S. Hochreiter and J. Schmidhuber. Long short-term memory.
Neural computation, 9(8), 1997. 1, 3
P. Hodosh, A. Young, M. Lai, and J. Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. In TACL,
2014. 6
H. Huang, Y. Lu, F. Zhang, and S. Sun. A multi-modal clustering method for web videos. In ISCTCS. 2013. 2
Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional
architecture for fast feature embedding. ACMMM, 2014. 2
A. Karpathy and L. Fei-Fei. Deep visual-semantic alignments for generating image descriptions. CVPR, 2015. 1
D. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7
R. Kiros, R. Salakhutdinov, and R. S. Zemel. Unifying
visual-semantic embeddings with multimodal neural language models. arXiv:1411.2539, 2014. 1
N. Krishnamoorthy, G. Malkarnenkar, R. J. Mooney,
K. Saenko, and S. Guadarrama. Generating natural-language
video descriptions using text-mined knowledge. In AAAI,
July 2013. 2
P. Kuznetsova, V. Ordonez, T. L. Berg, U. C. Hill, and
Y. Choi. Treetalk: Composition and compression of trees
for image descriptions. In TACL, 2014. 1
C.-Y. Lin. Rouge: A package for automatic evaluation of
summaries. In Text Summarization Branches Out: Proceedings of the ACL-04 Workshop, pages 7481, 2004. 5
T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollar, and C. L. Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6
J. Mao, W. Xu, Y. Yang, J. Wang, and A. L. Yuille. Deep
captioning with multimodal recurrent neural networks (mrnn). arXiv:1412.6632, 2014. 1

[24] J. Y. Ng, M. J. Hausknecht, S. Vijayanarasimhan, O. Vinyals,

R. Monga, and G. Toderici. Beyond short snippets: Deep
networks for video classification. CVPR, 2015. 3, 4
[25] P. Over, G. Awad, M. Michel, J. Fiscus, G. Sanders, B. Shaw,
A. F. Smeaton, and G. Queenot. TRECVID 2012 an
overview of the goals, tasks, data, evaluation mechanisms
and metrics. In Proceedings of TRECVID 2012, 2012. 2
[26] K. Papineni, S. Roukos, T. Ward, and W.-J. Zhu. Bleu: a
method for automatic evaluation of machine translation. In
ACL, 2002. 5
[27] A. Rohrbach, M. Rohrbach, and B. Schiele. The long-short
story of movie description. GCPR, 2015. 7
[28] A. Rohrbach, M. Rohrbach, N. Tandon, and B. Schiele. A
dataset for movie description. In CVPR, 2015. 1, 2, 5, 7
[29] M. Rohrbach, W. Qiu, I. Titov, S. Thater, M. Pinkal, and
B. Schiele. Translating video content to natural language
descriptions. In ICCV, 2013. 1
[30] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh,
S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein,
A. C. Berg, and L. Fei-Fei. ILSVRC, 2014. 4
[31] K. Simonyan and A. Zisserman. Two-stream convolutional
networks for action recognition in videos. In NIPS, 2014. 2
[32] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR,
abs/1409.1556, 2014. 4
[33] N. Srivastava, E. Mansimov, and R. Salakhutdinov. Unsupervised learning of video representations using LSTMs. ICML,
2015. 2
[34] I. Sutskever, O. Vinyals, and Q. V. Le. Sequence to sequence
learning with neural networks. In NIPS, 2014. 1, 2, 3
[35] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed,
D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich.
Going deeper with convolutions. CVPR, 2015. 6
[36] J. Thomason, S. Venugopalan, S. Guadarrama, K. Saenko,
and R. J. Mooney. Integrating language and vision to generate natural language descriptions of videos in the wild. In
COLING, 2014. 2, 6
[37] A. Torabi, C. Pal, H. Larochelle, and A. Courville. Using
descriptive video services to create a large data source for
video annotation research. arXiv:1503.01070v1, 2015. 2, 5
[38] R. Vedantam, C. L. Zitnick, and D. Parikh. CIDEr:
Consensus-based image description evaluation. CVPR,
2015. 5
[39] S. Venugopalan, H. Xu, J. Donahue, M. Rohrbach,
R. Mooney, and K. Saenko. Translating videos to natural
language using deep recurrent neural networks. In NAACL,
2015. 1, 2, 4, 5, 6, 7
[40] O. Vinyals, A. Toshev, S. Bengio, and D. Erhan. Show and
tell: A neural image caption generator. CVPR, 2015. 1, 2, 4
[41] H. Wang and C. Schmid. Action recognition with improved
trajectories. In ICCV, pages 35513558. IEEE, 2013. 2
[42] S. Wei, Y. Zhao, Z. Zhu, and N. Liu. Multimodal fusion for
video search reranking. TKDE, 2010. 2
[43] L. Yao, A. Torabi, K. Cho, N. Ballas, C. Pal, H. Larochelle,
and A. Courville. Describing videos by exploiting temporal
structure. arXiv:1502.08029v4, 2015. 1, 2, 4, 6, 7, 8
[44] W. Zaremba and I. Sutskever.
Learning to execute.
arXiv:1410.4615, 2014. 3