Lin2014 Chapter MicrosoftCOCOCommonObjectsInCo
Lin2014 Chapter MicrosoftCOCOCommonObjectsInCo
Tsung-Yi Lin1 , Michael Maire2 , Serge Belongie1 , James Hays3 , Pietro Perona2,
Deva Ramanan4 , Piotr Dollár5 , and C. Lawrence Zitnick5
1
Cornell
2
Caltech
3
Brown
4
UC Irvine
5
Microsoft Research
1 Introduction
One of the primary goals of computer vision is the understanding of visual scenes.
Scene understanding involves numerous tasks including recognizing what objects
are present, localizing the objects in 2D and 3D, determining the objects’ and
scene’s attributes, characterizing relationships between objects and providing a
semantic description of the scene. The current object classification and detection
datasets [1,2,3,4] help us explore the first challenges related to scene understand-
ing. For instance the ImageNet dataset [1], which contains an unprecedented
number of images, has recently enabled breakthroughs in both object classifi-
cation and detection research [5,6,7]. The community has also created datasets
containing object attributes [8], scene attributes [9], keypoints [10], and 3D scene
information [11]. This leads us to the obvious question: what datasets will best
continue our advance towards our ultimate goal of scene understanding?
We introduce a new large-scale dataset that addresses three core research
problems in scene understanding: detecting non-iconic views (or non-canonical
perspectives [12]) of objects, contextual reasoning between objects and the pre-
cise 2D localization of objects. For many categories of objects, there exists an
iconic view. For example, when performing a web-based image search for the
D. Fleet et al. (Eds.): ECCV 2014, Part V, LNCS 8693, pp. 740–755, 2014.
c Springer International Publishing Switzerland 2014
Microsoft COCO: Common Objects in Context 741
Fig. 1. While previous object recognition datasets have focused on (a) image classifi-
cation, (b) object bounding box localization or (c) semantic pixel-level segmentation,
we focus on (d) segmenting individual object instances. We introduce a large, richly-
annotated dataset comprised of images depicting complex everyday scenes of common
objects in their natural context
object category “bike,” the top-ranked retrieved examples appear in profile, un-
obstructed near the center of a neatly composed photo. We posit that current
recognition systems perform fairly well on iconic views, but struggle to recognize
objects otherwise – in the background, partially occluded, amid clutter [13] – re-
flecting the composition of actual everyday scenes. We verify this experimentally;
when evaluated on everyday scenes, models trained on our data perform better
than those trained with prior datasets. A challenge is finding natural images
that contain multiple objects. The identity of many objects can only be resolved
using context, due to small size or ambiguous appearance in the image. To push
research in contextual reasoning, images depicting scenes [3] rather than objects
in isolation are necessary. Finally, we argue that detailed spatial understanding
of object layout will be a core component of scene analysis. An object’s spa-
tial location can be defined coarsely using a bounding box [2] or with a precise
pixel-level segmentation [14,15,16]. As we demonstrate, to measure either kind
of localization performance it is essential for the dataset to have every instance
of every object category labeled and fully segmented. Our dataset is unique in
its annotation of instance-level segmentation masks, Fig. 1.
To create a large-scale dataset that accomplishes these three goals we em-
ployed a novel pipeline for gathering data with extensive use of Amazon Mechan-
ical Turk. First and most importantly, we harvested a large set of images con-
taining contextual relationships and non-iconic object views. We accomplished
this using a surprisingly simple yet effective technique that queries for pairs of
objects in conjunction with images retrieved via scene-based queries [17,3]. Next,
each image was labeled as containing particular object categories using a hierar-
chical labeling approach [18]. For each category found, the individual instances
were labeled, verified, and finally segmented. Given the inherent ambiguity of
labeling, each of these stages has numerous tradeoffs that we explored in detail.
The Microsoft Common Objects in COntext (MS COCO) dataset contains
91 common object categories with 82 of them having more than 5,000 labeled
instances, Fig. 6. In total the dataset has 2,500,000 labeled instances in 328,000
images. In contrast to the popular ImageNet dataset [1], COCO has fewer cate-
gories but more instances per category. This can aid in learning detailed object
models capable of precise 2D localization. The dataset is also significantly larger
in number of instances per category than the PASCAL VOC [2] and SUN [3]
datasets. Additionally, a critical distinction between our dataset and others is
742 T.-Y. Lin et al.
Fig. 2. Example of (a) iconic object images, (b) iconic scene images, and (c) non-iconic
images. In this work we focus on challenging non-iconic images.
the number of labeled instances per image which may aid in learning contex-
tual information, Fig. 5. MS COCO contains considerably more object instances
per image (7.7) as compared to ImageNet (3.0) and PASCAL (2.3). In contrast,
the SUN dataset, which contains significant contextual information, has over 17
objects and “stuff” per image but considerably fewer object instances overall.
An extended version of this work with additional details is available [19].
2 Related Work
Throughout the history of computer vision research datasets have played a crit-
ical role. They not only provide a means to train and evaluate algorithms, they
drive research in new and more challenging directions. The creation of ground
truth stereo and optical flow datasets [20,21] helped stimulate a flood of interest
in these areas. The early evolution of object recognition datasets [22,23,24] facil-
itated the direct comparison of hundreds of image recognition algorithms while
simultaneously pushing the field towards more complex problems. Recently, the
ImageNet dataset [1] containing millions of images has enabled breakthroughs
in both object classification and detection research using a new class of deep
learning algorithms [5,6,7].
Datasets related to object recognition can be roughly split into three groups:
those that primarily address object classification, object detection and semantic
scene labeling. We address each in turn.
Image Classification. The task of object classification requires binary labels
indicating whether objects are present in an image; see Fig. 1(a). Early datasets
of this type comprised images containing a single object with blank backgrounds,
such as the MNIST handwritten digits [25] or COIL household objects [26].
Caltech 101 [22] and Caltech 256 [23] marked the transition to more realistic
object images retrieved from the internet while also increasing the number of
object categories to 101 and 256, respectively. Popular datasets in the machine
learning community due to the larger number of training examples, CIFAR-10
and CIFAR-100 [27] offered 10 and 100 categories from a dataset of tiny 32 × 32
images [28]. While these datasets contained up to 60,000 images and hundreds
of categories, they still only captured a small fraction of our visual world.
Microsoft COCO: Common Objects in Context 743
Recently, ImageNet [1] made a striking departure from the incremental in-
crease in dataset sizes. They proposed the creation of a dataset containing 22k
categories with 500-1000 images each. Unlike previous datasets containing entry-
level categories [29], such as “dog” or “chair,” like [28], ImageNet used the Word-
Net Hierarchy [30] to obtain both entry-level and fine-grained [31] categories.
Currently, the ImageNet dataset contains over 14 million labeled images and
has enabled significant advances in image classification [5,6,7].
Object Detection. Detecting an object entails both stating that an object
belonging to a specified class is present, and localizing it in the image. The
location of an object is typically represented by a bounding box, Fig. 1(b). Early
algorithms focused on face detection [32] using various ad hoc datasets. Later,
more realistic and challenging face detection datasets were created [33]. Another
popular challenge is the detection of pedestrians for which several datasets have
been created [24,4]. The Caltech Pedestrian Dataset [4] contains 350,000 labeled
instances with bounding boxes.
For the detection of basic object categories, a multi-year effort from 2005 to
2012 was devoted to the creation and maintenance of a series of benchmark
datasets that were widely adopted. The PASCAL VOC [2] datasets contained
20 object categories spread over 11,000 images. Over 27,000 object instance
bounding boxes were labeled, of which almost 7,000 had detailed segmentations.
Recently, a detection challenge has been created from 200 object categories using
a subset of 400,000 images from ImageNet [34]. An impressive 350,000 objects
have been labeled using bounding boxes.
Since the detection of many objects such as sunglasses, cellphones or chairs
is highly dependent on contextual information, it is important that detection
datasets contain objects in their natural environments. In our dataset we strive
to collect images rich in contextual information. The use of bounding boxes also
limits the accuracy for which detection algorithms may be evaluated. We propose
the use of fully segmented instances to enable more accurate detector evaluation.
Semantic Scene Labeling. The task of labeling semantic objects in a scene
requires that each pixel of an image be labeled as belonging to a category, such as
sky, chair, floor, street, etc. In contrast to the detection task, individual instances
of objects do not need to be segmented, Fig. 1(c). This enables the labeling of
objects for which individual instances are hard to define, such as grass, streets,
or walls. Datasets exist for both indoor [11] and outdoor [35,14] scenes. Some
datasets also include depth information [11]. Similar to semantic scene labeling,
our goal is to measure the pixel-wise accuracy of object labels. However, we also
aim to distinguish between individual instances of an object, which requires a
solid understanding of each object’s extent.
A novel dataset that combines many of the properties of both object de-
tection and semantic scene labeling datasets is the SUN dataset [3] for scene
understanding. SUN contains 908 scene categories from the WordNet dictionary
[30] with segmented objects. The 3,819 object categories span those common
to object detection datasets (person, chair, car) and to semantic scene labeling
744 T.-Y. Lin et al.
(wall, sky, floor). Since the dataset was collected by finding images depicting
various scene types, the number of instances per object category exhibits the
long tail phenomenon. That is, a few categories have a large number of instances
(wall: 20,213, window: 16,080, chair: 7,971) while most have a relatively modest
number of instances (boat: 349, airplane: 179, floor lamp: 276). In our dataset,
we ensure that each object category has a significant number of instances, Fig. 5.
Other Vision Datasets. Datasets have spurred the advancement of numer-
ous fields in computer vision. Some notable datasets include the Middlebury
datasets for stereo vision [20], multi-view stereo [36] and optical flow [21]. The
Berkeley Segmentation Data Set (BSDS500) [37] has been used extensively to
evaluate both segmentation and edge detection algorithms. Datasets have also
been created to recognize both scene [9] and object attributes [8,38]. Indeed,
numerous areas of vision have benefited from challenging datasets that helped
catalyze progress.
3 Image Collection
We next describe how the object categories and candidate images are selected.
Common Object Categories. The selection of object categories is a non-
trivial exercise. The categories must form a representative set of all categories,
be relevant to practical applications and occur with high enough frequency to
enable the collection of a large dataset. Other important decisions are whether
to include both “thing” and “stuff” categories [39] and whether fine-grained
[31,1] and object-part categories should be included. “Thing” categories include
objects for which individual instances may be easily labeled (person, chair, car)
where “stuff” categories include materials and objects with no clear boundaries
(sky, street, grass). Since we are primarily interested in precise localization of
object instances, we decided to only include “thing” categories and not “stuff.”
However, since “stuff” categories can provide significant contextual information,
we believe the future labeling of “stuff” categories would be beneficial.
The specificity of object categories can vary significantly. For instance, a dog
could be a member of the “mammal”, “dog”, or “German shepherd” categories.
To enable the practical collection of a significant number of instances per cate-
gory, we chose to limit our dataset to entry-level categories, i.e. category labels
that are commonly used by humans when describing objects (dog, chair, per-
son). It is also possible that some object categories may be parts of other object
categories. For instance, a face may be part of a person. We anticipate the inclu-
sion of object-part categories (face, hands, wheels) would be beneficial for many
real-world applications.
We used several sources to collect entry-level object categories of “things.” We
first compiled a list of categories by combining categories from PASCAL VOC
[2] and a subset of the 1200 most frequently used words that denote visually
identifiable objects [40]. To further augment our set of candidate categories,
several children ranging in ages from 4 to 8 were asked to name every object
Microsoft COCO: Common Objects in Context 745
they see in indoor and outdoor environments. The final 271 candidates may be
found in [19]. Finally, the co-authors voted on a 1 to 5 scale for each category
taking into account how commonly they occur, their usefulness for practical
applications, and their diversity relative to other categories. The final selection of
categories attempts to pick categories with high votes, while keeping the number
of categories per super-category (animals, vehicles, furniture, etc.) balanced.
Categories for which obtaining a large number of instances (greater than 5,000)
was difficult were also removed. To ensure backwards compatibility all categories
from PASCAL VOC [2] are also included. Our final list of 91 proposed categories
is in Fig. 5(a).
Non-iconic Image Collection. Given the list of object categories, our next
goal was to collect a set of candidate images. We may roughly group images
into three types, Fig. 2: iconic-object images [41], iconic-scene images [3] and
non-iconic images. Typical iconic-object images have a single large object in a
canonical perspective centered in the image, Fig. 2(a). Iconic-scene images are
shot from canonical viewpoints and commonly lack people, Fig. 2(b). Iconic im-
ages have the benefit that they may be easily found by directly searching for
specific categories using Google or Bing image search. While iconic images gen-
erally provide high quality object instances, they can lack important contextual
information and non-canonical viewpoints.
Our goal was to collect a dataset such that a majority of images are non-
iconic, Fig. 2(c). It has been shown that datasets containing more non-iconic
images are better at generalizing [42]. We collected non-iconic images using two
strategies. First as popularized by PASCAL VOC [2], we collected images from
Flickr which tends to have fewer iconic images. Flickr contains photos uploaded
by amateur photographers with searchable metadata and keywords. Second, we
did not search for object categories in isolation. A search for “dog” will tend to
return iconic images of large, centered dogs. However, if we searched for pairwise
combinations of object categories, such as “dog + car” we found many more non-
iconic images. Surprisingly, these images typically do not just contain the two
categories specified in the search, but numerous other categories as well. To
further supplement our dataset we also searched for scene/object category pairs,
see [19]. We downloaded at most 5 photos taken by a single photographer within
a short time window. In the rare cases in which enough images could not be
found, we searched for single categories and performed an explicit filtering stage
to remove iconic images. The result is a collection of 328,000 images with rich
contextual relationships between objects as shown in Figs. 2(c) and 6.
4 Image Annotation
We next describe how we annotated our image collection. Due to our desire to
label over 2.5 million category instances, the design of a cost efficient yet high
quality annotation pipeline was critical. The annotation pipeline is outlined in
Fig. 3. For all crowdsourcing tasks we used workers on Amazon’s Mechanical
Turk (AMT). Examples of our user interfaces can be found in [19].
746 T.-Y. Lin et al.
Fig. 3. Our image annotation pipeline is split into 3 primary worker tasks: (a) Labeling
the categories present in the image, (b) locating and marking all instances of the labeled
categories, and (c) segmenting each object instance.
1 16
precision
0.6
8
0.4
4
0.2
0 0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
recall
(a) (b)
Fig. 4. (a) Precision and recall of experts (red) and the majority vote of AMT workers
(blue). Note that the aggregate of 3 workers has better or similar recall to most experts.
(b) illustrates the precision and recall of workers, with color indicating how many jobs
they completed. For details and definition of ground truth for each plot see text.
results obtained by taking the union of one to ten AMT workers. For this task
precision is of less importance since false positives will be removed at later stages,
where adding false negatives is much more difficult. Fig. 4(a) shows that 5 AMT
workers, the same number as was used to collect our labels, achieves the same
recall as most of the expert workers. Note that the expert labelers achieved
between 65% and 80% recall. These low values of recall are due to our liberal
definition of a category being present. If only one expert labels an object category
as being present, we assume the category is indeed present. However, the presence
of many categories is often ambiguous. Upon closer inspection, we find recall
values of 70% to 75% are generally sufficient to capture the non-ambiguous
categories. Fig. 4(b) shows the precision and recall of our workers on category
labeling. Unlike in Fig. 4(a), the ground truth labels were now estimated using a
majority vote. The color indicates the number of jobs completed by each worker.
Notice that workers who complete more hits have generally higher precision and
recall. All jobs from workers below the black line were rejected.
5 Dataset Statistics
Fig. 5. (a) Number of annotated instances per category for MS COCO and PASCAL
VOC. (b,c) Number of annotated categories and annotated instances, respectively, per
image for MS COCO, ImageNet Detection, PASCAL VOC and SUN (average number
of categories and instances are shown in parentheses). (d) Number of categories vs. the
number of instances per category for a number of popular object recognition datasets.
(e) The distribution of instance sizes for the MS COCO, ImageNet Detection, PASCAL
VOC and SUN datasets.
over 60% of images containing a single object category in ImageNet and PAS-
CAL VOC. As expected, the SUN dataset has the most contextual information
since it is scene-based.
Finally, we analyze the average size of objects in the datasets. Generally
smaller objects are harder to recognize and require more contextual reasoning
to recognize. As shown in Fig. 5(e), the average sizes of objects is smaller for
both MS COCO and SUN.
750 T.-Y. Lin et al.
6 Algorithmic Analysis
models tested on both the MS COCO and PASCAL datasets, see Table 1. We eval-
uate two different models. DPMv5-P: the latest implementation of [44] (release 5
[45]) trained on PASCAL VOC 2012. DPMv5-C: the same implementation
trained on COCO (5000 positive and 10000 negative images). We use the default
parameter settings for training COCO models.
If we compare the average performance of DPMv5-P on PASCAL VOC and
MS COCO, we find that average performance on MS COCO drops by nearly a
factor of 2, suggesting that MS COCO does include more difficult (non-iconic)
images of objects that are partially occluded, amid clutter, etc. We notice a
similar drop in performance for the model trained on MS COCO (DPMv5-C).
The effect on detection performance of training on PASCAL VOC or MS
COCO may be analyzed by comparing DPMv5-P and DPMv5-C. They use
the same implementation with different sources of training data. Table 1 shows
DPMv5-C still outperforms DPMv5-P in 6 out of 20 categories when testing on
PASCAL VOC. In some categories (e.g., dog, cat, people), models trained on
MS COCO perform worse, while on others (e.g., bus, tv, horse), models trained
on our data are better.
Consistent with past observations [46], we find that including difficult (non-
iconic) images during training may not always help. Such examples may act as
noise and pollute the learned model if the model is not rich enough to capture
such appearance variability. Our dataset allows for the exploration of such issues.
Torralba and Efros [42] proposed a metric to measure cross-dataset gener-
alization which computes the ‘performance drop’ for models that train on one
dataset and test on another. The performance difference of the DPMv5-P mod-
els across the two datasets is 12.7 AP while the DPMv5-C models only have
7.7 AP difference. Moreover, overall performance is much lower on MS COCO.
These observations support two hypotheses: 1) MS COCO is significantly more
difficult than PASCAL VOC and 2) models trained on MS COCO can generalize
better to easier datasets such as PASCAL VOC given more training data. To
gain insight into the differences between the datasets, see [19] for visualizations
of person and chair examples from the two datasets.
752 T.-Y. Lin et al.
0.8
0.8
0.7
Segmentation Overlap
0.7
Groundtruth
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
Predicted
0.2
0.2
0.1
0.1
0
0.5 0.6 0.7 0.8 0.9 1
Bounding Box Overlap 0
fa
p e
s
ni ow
bi e
e
at
tte on
r
sh t
p
m in
r
rd
g
t
ot se
n
ca
ca
ai
ito
bu
bl
k
an
cl
ee
ttl
do
tra
so
bo
la
bi
bi
r
po ers
ch
c
cy
ta
bo
on
ho
dp
pl
or
ng
ro
ae
tv
m
di
Fig. 9. A predicted segmentation might not recover object detail even though detection
and groundtruth bounding boxes overlap well (left). Sampling from the person category
illustrates that on a per-instance basis, predicting segmentation from top-down pro-
jection of DPM part masks is difficult even for correct detections (center). Averaging
over instances for each of the PASCAL VOC categories on our dataset demonstrates
that it presents a challenge for object segmentation algorithms (right).
7 Discussion
We described a new dataset for detecting and segmenting objects found in ev-
eryday life in their natural environments. Utilizing around 60,000 worker hours,
a vast collection of category instances was gathered, annotated and organized to
drive the advancement of object detection and segmentation algorithms. Empha-
sis was placed on finding non-iconic images of objects in natural environments
and varied viewpoints. Dataset statistics indicate the images contain rich con-
textual information with many objects present per image.
There are several promising directions for future annotations on our dataset.
We currently only label “things”, but labeling “stuff” may also provide sig-
nificant contextual information that may be useful for detection. Many object
detection algorithms benefit from additional annotations, such as the amount an
instance is occluded [4] or the location of keypoints on the object [10]. Finally,
our dataset could provide a good benchmark for other types of labels, including
scene types [3], attributes [9,8] and full sentence written descriptions [51].
To download and learn more about MS COCO please see the project website1 .
Additional details are presented in an extended version of this work [19]. MS
COCO will evolve and grow over time; up to date information is available online.
Acknowledgments. Funding for all crowd worker tasks was provided by Mi-
crosoft. P.P. and D.R. were supported by ONR MURI Grant N00014-10-1-0933.
We would like to thank all members of the community who provided valuable
feedback throughout the process of defining and collecting the dataset.
1
http://mscoco.org/
754 T.-Y. Lin et al.
References
1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A Large-
Scale Hierarchical Image Database. In: CVPR (2009)
2. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The
PASCAL visual object classes (VOC) challenge. IJCV 88(2), 303–338 (2010)
3. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: SUN database: Large-
scale scene recognition from abbey to zoo. In: CVPR (2010)
4. Dollár, P., Wojek, C., Schiele, B., Perona, P.: Pedestrian detection: An evaluation
of the state of the art. PAMI 34 (2012)
5. Krizhevsky, A., Sutskever, I., Hinton, G.: ImageNet classification with deep con-
volutional neural networks. In: NIPS (2012)
6. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accu-
rate object detection and semantic segmentation. In: CVPR (2014)
7. Sermanet, P., Eigen, D., Zhang, S., Mathieu, M., Fergus, R., LeCun, Y.: OverFeat:
Integrated recognition, localization and detection using convolutional networks. In:
ICLR (April 2014)
8. Farhadi, A., Endres, I., Hoiem, D., Forsyth, D.: Describing objects by their at-
tributes. In: CVPR (2009)
9. Patterson, G., Hays, J.: SUN attribute database: Discovering, annotating, and
recognizing scene attributes. In: CVPR (2012)
10. Bourdev, L., Malik, J.: Poselets: Body part detectors trained using 3D human pose
annotations. In: ICCV (2009)
11. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support
inference from RGBD images. In: Fitzgibbon, A., Lazebnik, S., Perona, P., Sato, Y.,
Schmid, C. (eds.) ECCV 2012, Part V. LNCS, vol. 7576, pp. 746–760. Springer, Hei-
delberg (2012)
12. Palmer, S., Rosch, E., Chase, P.: Canonical perspective and the perception of
objects. Attention and Performance IX 1, 4 (1981)
13. Hoiem, D., Chodpathumwan, Y., Dai, Q.: Diagnosing error in object detectors. In:
Fitzgibbon, A., Lazebnik, S., Perona, P., Sato, Y., Schmid, C. (eds.) ECCV 2012,
Part III. LNCS, vol. 7574, pp. 340–353. Springer, Heidelberg (2012)
14. Brostow, G., Fauqueur, J., Cipolla, R.: Semantic object classes in video: A high-
definition ground truth database. PRL 30(2), 88–97 (2009)
15. Russell, B., Torralba, A., Murphy, K., Freeman, W.: LabelMe: a database and
web-based tool for image annotation. IJCV 77(1-3), 157–173 (2008)
16. Bell, S., Upchurch, P., Snavely, N., Bala, K.: OpenSurfaces: A richly annotated
catalog of surface appearance. SIGGRAPH 32(4) (2013)
17. Ordonez, V., Kulkarni, G., Berg, T.: Im2text: Describing images using 1 million
captioned photographs. In: NIPS (2011)
18. Deng, J., Russakovsky, O., Krause, J., Bernstein, M., Berg, A., Fei-Fei, L.: Scalable
multi-label annotation. In: CHI (2014)
19. Lin, T., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P.,
Zitnick, C.L.: Microsoft COCO: Common objects in context. CoRR abs/1405.0312
(2014)
20. Scharstein, D., Szeliski, R.: A taxonomy and evaluation of dense two-frame stereo
correspondence algorithms. IJCV 47(1-3), 7–42 (2002)
21. Baker, S., Scharstein, D., Lewis, J., Roth, S., Black, M., Szeliski, R.: A database
and evaluation methodology for optical flow. IJCV 92(1), 1–31 (2011)
22. Fei-Fei, L., Fergus, R., Perona, P.: Learning generative visual models from few
training examples: An incremental bayesian approach tested on 101 object cate-
gories. In: CVPR Workshop of Generative Model Based Vision, WGMBV (2004)
Microsoft COCO: Common Objects in Context 755
23. Griffin, G., Holub, A., Perona, P.: Caltech-256 object category dataset. Technical
Report 7694, California Institute of Technology (2007)
24. Dalal, N., Triggs, B.: Histograms of oriented gradients for human detection. In:
CVPR (2005)
25. Lecun, Y., Cortes, C.: The MNIST database of handwritten digits (1998)
26. Nene, S.A., Nayar, S.K., Murase, H.: Columbia object image library (coil-20). Tech-
nical report, Columbia Universty (1996)
27. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images.
Computer Science Department, University of Toronto, Tech. Rep. (2009)
28. Torralba, A., Fergus, R., Freeman, W.T.: 80 million tiny images: A large data set
for nonparametric object and scene recognition. PAMI 30(11), 1958–1970 (2008)
29. Ordonez, V., Deng, J., Choi, Y., Berg, A., Berg, T.: From large scale image cate-
gorization to entry-level categories. In: ICCV (2013)
30. Fellbaum, C.: WordNet: An electronic lexical database. Blackwell Books (1998)
31. Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., Perona, P.:
Caltech-UCSD Birds 200. Technical Report CNS-TR-201, Caltech. (2010)
32. Hjelmås, E., Low, B.: Face detection: A survey. CVIU 83(3), 236–274 (2001)
33. Huang, G.B., Ramesh, M., Berg, T., Learned-Miller, E.: Labeled faces in the wild.
Technical Report 07-49, University of Massachusetts, Amherst (October 2007)
34. Russakovsky, O., Deng, J., Huang, Z., Berg, A., Fei-Fei, L.: Detecting avocados to
zucchinis: what have we done, and where are we going? In: ICCV (2013)
35. Shotton, J., Winn, J., Rother, C., Criminisi, A.: TextonBoost for image understand-
ing: Multi-class object recognition and segmentation by jointly modeling texture,
layout, and context. IJCV 81(1), 2–23 (2009)
36. Seitz, S.M., Curless, B., Diebel, J., Scharstein, D., Szeliski, R.: A comparison and
evaluation of multi-view stereo reconstruction algorithms. In: CVPR (2006)
37. Arbelaez, P., Maire, M., Fowlkes, C., Malik, J.: Contour detection and hierarchical
image segmentation. PAMI 33(5), 898–916 (2011)
38. Lampert, C., Nickisch, H., Harmeling, S.: Learning to detect unseen object classes
by between-class attribute transfer. In: CVPR (2009)
39. Heitz, G., Koller, D.: Learning spatial context: Using stuff to find things. In:
Forsyth, D., Torr, P., Zisserman, A. (eds.) ECCV 2008, Part I. LNCS, vol. 5302,
pp. 30–43. Springer, Heidelberg (2008)
40. Sitton, R.: Spelling Sourcebook. Egger Publishing (1996)
41. Berg, T., Berg, A.: Finding iconic images. In: CVPR (2009)
42. Torralba, A., Efros, A.: Unbiased look at dataset bias. In: CVPR (2011)
43. Douze, M., Jégou, H., Sandhawalia, H., Amsaleg, L., Schmid, C.: Evaluation of
gist descriptors for web-scale image search. In: CIVR (2009)
44. Felzenszwalb, P., Girshick, R., McAllester, D., Ramanan, D.: Object detection with
discriminatively trained part-based models. PAMI 32(9), 1627–1645 (2010)
45. Girshick, R., Felzenszwalb, P., McAllester, D.: Discriminatively trained deformable
part models, release 5. PAMI (2012)
46. Zhu, X., Vondrick, C., Ramanan, D., Fowlkes, C.: Do we need more training data
or better models for object detection? In: BMVC (2012)
47. Brox, T., Bourdev, L., Maji, S., Malik, J.: Object segmentation by alignment of
poselet activations to image contours. In: CVPR (2011)
48. Yang, Y., Hallman, S., Ramanan, D., Fowlkes, C.: Layered object models for image
segmentation. PAMI 34(9), 1731–1743 (2012)
49. Ramanan, D.: Using segmentation to verify object hypotheses. In: CVPR (2007)
50. Dai, Q., Hoiem, D.: Learning to localize detected objects. In: CVPR (2012)
51. Rashtchian, C., Young, P., Hodosh, M., Hockenmaier, J.: Collecting image anno-
tations using Amazon’s Mechanical Turk. In: NAACL Workshop (2010)