Attached is a paper on CNN. It must be the basis of your research paper. You must analyze the proposal, explain the experiment and critic the result. You must research any term that is not well defined in the paper using other publish literatures. 

Large-scale Video Classification with Convolutional Neural Networks

Andrej Karpathy1,2 George Toderici1 Sanketh Shetty1

karpathy@cs.stanford.edu gtoderici@google.com sanketh@google.com

Thomas Leung1 Rahul Sukthankar1 Li Fei-Fei2

leungt@google.com sukthankar@google.com feifeili@cs.stanford.edu
1Google Research 2Computer Science Department, Stanford University
http://cs.stanford.edu/people/karpathy/deepvideo

Abstract

Convolutional Neural Networks (CNNs) have been es-
tablished as a powerful class of models for image recog-
nition problems. Encouraged by these results, we pro-
vide an extensive empirical evaluation of CNNs on large-
scale video classification using a new dataset of 1 million
YouTube videos belonging to 487 classes. We study mul-
tiple approaches for extending the connectivity of a CNN
in time domain to take advantage of local spatio-temporal
information and suggest a multiresolution, foveated archi-
tecture as a promising way of speeding up the training.
Our best spatio-temporal networks display significant per-
formance improvements compared to strong feature-based
baselines (55.3% to 63.9%), but only a surprisingly mod-
est improvement compared to single-frame models (59.3%
to 60.9%). We further study the generalization performance
of our best model by retraining the top layers on the UCF-
101 Action Recognition dataset and observe significant per-
formance improvements compared to the UCF-101 baseline
model (63.3% up from 43.9%).

1. Introduction
Images and videos have become ubiquitous on the in-

ternet, which has encouraged the development of algo-
rithms that can analyze their semantic content for vari-
ous applications, including search and summarization. Re-
cently, Convolutional Neural Networks (CNNs) [15] have
been demonstrated as an effective class of models for un-
derstanding image content, giving state-of-the-art results
on image recognition, segmentation, detection and retrieval
[11, 3, 2, 20, 9, 18]. The key enabling factors behind these
results were techniques for scaling up the networks to tens
of millions of parameters and massive labeled datasets that
can support the learning process. Under these conditions,
CNNs have been shown to learn powerful and interpretable

image features [28]. Encouraged by positive results in do-
main of images, we study the performance of CNNs in
large-scale video classification, where the networks have
access to not only the appearance information present in
single, static images, but also their complex temporal evolu-
tion. There are several challenges to extending and applying
CNNs in this setting.

From a practical standpoint, there are currently no video
classification benchmarks that match the scale and variety
of existing image datasets because videos are significantly
more difficult to collect, annotate and store. To obtain suffi-
cient amount of data needed to train our CNN architectures,
we collected a new Sports-1M dataset, which consists of 1
million YouTube videos belonging to a taxonomy of 487
classes of sports. We make Sports-1M available to the re-
search community to support future work in this area.

From a modeling perspective, we are interested in an-
swering the following questions: what temporal connectiv-
ity pattern in a CNN architecture is best at taking advantage
of local motion information present in the video? How does
the additional motion information influence the predictions
of a CNN and how much does it improve performance over-
all? We examine these questions empirically by evaluating
multiple CNN architectures that each take a different ap-
proach to combining information across the time domain.

From a computational perspective, CNNs require exten-
sively long periods of training time to effectively optimize
the millions of parameters that parametrize the model. This
difficulty is further compounded when extending the con-
nectivity of the architecture in time because the network
must process not just one image but several frames of video
at a time. To mitigate this issue, we show that an effec-
tive approach to speeding up the runtime performance of
CNNs is to modify the architecture to contain two separate
streams of processing: a context stream that learns features
on low-resolution frames and a high-resolution fovea stream
that only operates on the middle portion of the frame. We

http://cs.stanford.edu/people/karpathy/deepvideo

observe a 2-4x increase in runtime performance of the net-
work due to the reduced dimensionality of the input, while
retaining the classification accuracy.

Finally, a natural question that arises is whether features
learned on the Sports-1M dataset are generic enough to
generalize to a different, smaller dataset. We investigate
the transfer learning problem empirically, achieving sig-
nificantly better performance (65.4%, up from 41.3%) on
UCF-101 by re-purposing low-level features learned on the
Sports-1M dataset than by training the entire network on
UCF-101 alone. Furthermore, since only some classes in
UCF-101 are related to sports, we can quantify the relative
improvements of the transfer learning in both settings.

Our contributions can be summarized as follows:

• We provide extensive experimental evaluation of mul-
tiple approaches for extending CNNs into video clas-
sification on a large-scale dataset of 1 million videos
with 487 categories (which we release as Sports-1M
dataset) and report significant gains in performance
over strong feature-based baselines.

• We highlight an architecture that processes input at two
spatial resolutions – a low-resolution context stream
and a high-resolution fovea stream – as a promising
way of improving the runtime performance of CNNs
at no cost in accuracy.

• We apply our networks to the UCF-101 dataset and re-
port significant improvement over feature-based state-
of-the-art results and baselines established by training
networks on UCF-101 alone.

2. Related Work
The standard approach to video classification [26, 16,

21, 17] involves three major stages: First, local visual fea-
tures that describe a region of the video are extracted ei-
ther densely [25] or at a sparse set of interest points [12, 8].
Next, the features get combined into a fixed-sized video-
level description. One popular approach is to quantize all
features using a learned k-means dictionary and accumulate
the visual words over the duration of the video into his-
tograms of varying spatio-temporal positions and extents
[13]. Lastly, a classifier (such as an SVM) is trained on
the resulting ”bag of words” representation to distinguish
among the visual classes of interest.

Convolutional Neural Networks [15] are a biologically-
inspired class of deep learning models that replace all three
stages with a single neural network that is trained end to
end from raw pixel values to classifier outputs. The spa-
tial structure of images is explicitly taken advantage of for
regularization through restricted connectivity between lay-
ers (local filters), parameter sharing (convolutions) and spe-
cial local invariance-building neurons (max pooling). Thus,
these architectures effectively shift the required engineer-

ing from feature design and accumulation strategies to de-
sign of the network connectivity structure and hyperparam-
eter choices. Due to computational constraints, CNNs have
until recently been applied to relatively small scale image
recognition problems (on datasets such as MNIST, CIFAR-
10/100, NORB, and Caltech-101/256), but improvements
on GPU hardware have enabled CNNs to scale to networks
of millions of parameters, which has in turn led to signif-
icant improvements in image classification[11], object de-
tection [20, 9], scene labeling [3], indoor segmentation [4]
and house number digit classification [19]. Additionally,
features learned by large networks trained on ImageNet
[7] have been shown to yield state-of-the-art performance
across many standard image recognition datasets when clas-
sified with an SVM, even with no fine-tuning [18].

Compared to image data domains, there is relatively lit-
tle work on applying CNNs to video classification. Since
all successful applications of CNNs in image domains share
the availability of a large training set, we speculate that this
is partly attributable to lack of large-scale video classifica-
tion benchmarks. In particular, commonly used datasets
(KTH, Weizmann, UCF Sports, IXMAS, Hollywood 2,
UCF-50) only contain up to few thousand clips and up to
few dozen classes. Even the largest available datasets such
as CCV (9,317 videos and 20 classes) and the recently in-
troduced UCF-101[22] (13,320 videos and 101 classes) are
still dwarfed by available image datasets in the number of
instances and their variety [7]. Despite these limitations,
some extensions of CNNs into the video domain have been
explored. [1] and [10] extend an image CNN to video
domains by treating space and time as equivalent dimen-
sions of the input and perform convolutions in both time
and space. We consider these extensions as only one of the
possible generalizations in this work. Unsupervised learn-
ing schemes for training spatio-temporal features have also
been developed, based on Convolutional Gated Restricted
Boltzmann Machines [23] and Independent Subspace Anal-
ysis [14]. In contrast, our models are trained end to end
fully supervised.

3. Models
Unlike images which can be cropped and rescaled to a

fixed size, videos vary widely in temporal extent and can-
not be easily processed with a fixed-sized architecture. In
this work we treat every video as a bag of short, fixed-sized
clips. Since each clip contains several contiguous frames
in time, we can extend the connectivity of the network in
time dimension to learn spatio-temporal features. There are
multiple options for the precise details of the extended con-
nectivity and we describe three broad connectivity pattern
categories (Early Fusion, Late Fusion and Slow Fusion) be-
low. Afterwards, we describe a multiresolution architecture
for addressing the computational efficiency.

Figure 1: Explored approaches for fusing information over
temporal dimension through the network. Red, green and
blue boxes indicate convolutional, normalization and pool-
ing layers respectively. In the Slow Fusion model, the de-
picted columns share parameters.

3.1. Time Information Fusion in CNNs

We investigate several approaches to fusing information
across temporal domain (Figure 1): the fusion can be done
early in the network by modifying the first layer convolu-
tional filters to extend in time, or it can be done late by
placing two separate single-frame networks some distance
in time apart and fusing their outputs later in the process-
ing. We first describe a baseline single-frame CNN and then
discuss its extensions in time according to different types of
fusion.

Single-frame. We use a single-frame baseline architec-
ture to understand the contribution of static appearance to
the classification accuracy. This network is similar to the
ImageNet challenge winning model [11], but accepts in-
puts of size 170 × 170 × 3 pixels instead of the original
224 × 224 × 3. Using shorthand notation, the full architec-
ture is C(96,11,3)-N-P -C(256,5,1)-N-P -C(384,3,1)-
C(384,3,1)-C(256,3,1)-P -FC(4096)-FC(4096), where
C(d,f,s) indicates a convolutional layer with d filters of
spatial size f ×f, applied to the input with stride s. FC(n)
is a fully connected layer with n nodes. All pooling layers P
pool spatially in non-overlapping 2 × 2 regions and all nor-
malization layers N are defined as described in Krizhevsky
et al. [11] and use the same parameters: k = 2,n = 5,α =
10−4,β = 0.5. The final layer is connected to a softmax
classifier with dense connections.

Early Fusion. The Early Fusion extension combines in-
formation across an entire time window immediately on the
pixel level. This is implemented by modifying the filters on
the first convolutional layer in the single-frame model by
extending them to be of size 11 × 11 × 3 × T pixels, where
T is some temporal extent (we use T = 10, or approxi-
mately a third of a second). The early and direct connectiv-
ity to pixel data allows the network to precisely detect local
motion direction and speed.

Late Fusion. The Late Fusion model places two sepa-
rate single-frame networks (as described above, up to last
convolutional layer C(256,3,1) with shared parameters a
distance of 15 frames apart and then merges the two streams

in the first fully connected layer. Therefore, neither single-
frame tower alone can detect any motion, but the first fully
connected layer can compute global motion characteristics
by comparing outputs of both towers.

Slow Fusion. The Slow Fusion model is a balanced
mix between the two approaches that slowly fuses temporal
information throughout the network such that higher lay-
ers get access to progressively more global information in
both spatial and temporal dimensions. This is implemented
by extending the connectivity of all convolutional layers
in time and carrying out temporal convolutions in addition
to spatial convolutions to compute activations, as seen in
[1, 10]. In the model we use, the first convolutional layer is
extended to apply every filter of temporal extent T = 4 on
an input clip of 10 frames through valid convolution with
stride 2 and produces 4 responses in time. The second and
third layers above iterate this process with filters of tempo-
ral extent T = 2 and stride 2. Thus, the third convolutional
layer has access to information across all 10 input frames.

3.2. Multiresolution CNNs

Since CNNs normally take on orders of weeks to train on
large-scale datasets even on the fastest available GPUs, the
runtime performance is a critical component to our ability
to experiment with different architecture and hyperparame-
ter settings. This motivates approaches for speeding up the
models while still retaining their performance. There are
multiple fronts to these endeavors, including improvements
in hardware, weight quantization schemes, better optimiza-
tion algorithms and initialization strategies, but in this work
we focus on changes in the architecture that enable faster
running times without sacrificing performance.

One approach to speeding up the networks is to reduce
the number of layers and neurons in each layer, but simi-
lar to [28] we found that this consistently lowers the per-
formance. Instead of reducing the size of the network, we
conducted further experiments on training with images of
lower resolution. However, while this improved the run-
ning time of the network, the high-frequency detail in the
images proved critical to achieving good accuracy.

Fovea and context streams. The proposed multiresolu-
tion architecture aims to strike a compromise by having two
separate streams of processing over two spatial resolutions
(Figure 2). A 178 × 178 frame video clip forms an input
to the network. The context stream receives the downsam-
pled frames at half the original spatial resolution (89 × 89
pixels), while the fovea stream receives the center 89 × 89
region at the original resolution. In this way, the the total
input dimensionality is halved. Notably, this design takes
advantage of the camera bias present in many online videos,
since the object of interest often occupies the center region.

Architecture changes. Both streams are processed by
identical network as the full frame models, but starting at

Figure 2: Multiresolution CNN architecture. Input frames
are fed into two separate streams of processing: a con-
text stream that models low-resolution image and a fovea
stream that processes high-resolution center crop. Both
streams consist of alternating convolution (red), normaliza-
tion (green) and pooling (blue) layers. Both streams con-
verge to two fully connected layers (yellow).

89 × 89 clips of video. Since the input is only of half the
spatial size as the full-frame models, we take out the last
pooling layer to ensure that both streams still terminate in a
layer of size 7×7×256. The activations from both streams
are concatenated and fed into the first fully connected layer
with dense connections.

3.3. Learning
Optimization. We use Downpour Stochastic Gradient

Descent [6] to optimize our models across a computing
cluster. The number of replicas for each model varies be-
tween 10 and 50 and every model is further split across 4
to 32 partitions. We use mini-batches of 32 examples, mo-
mentum of 0.9 and weight decay of 0.0005. All models are
initialized with learning rates of 1e−3 and this value is fur-
ther reduced by hand whenever the validation error stops
improving.

Data augmentation and preprocessing. Following
[11], we take advantage of data augmentation to reduce the
effects of overfitting. Before presenting an example to a net-
work, we preprocess all images by first cropping to center
region, resizing them to 200 × 200 pixels, randomly sam-
pling a 170 × 170 region, and finally randomly flipping the
images horizontally with 50% probability. These prepro-
cessing steps are applied consistently to all frames that are
part of the same clip. As a last step of preprocessing we sub-
tract a constant value of 117 from raw pixel values, which
is the approximate value of the mean of all pixels in our
images.

4. Results
We first present results on our Sports-1M dataset and

qualitatively analyze the learned features and network pre-

dictions. We then describe our transfer learning experi-
ments on UCF-101.

4.1. Experiments on Sports-1M

Dataset. The Sports-1M dataset consists of 1 million
YouTube videos annotated with 487 classes. The classes
are arranged in a manually-curated taxonomy that contains
internal nodes such as Aquatic Sports, Team Sports, Winter
Sports, Ball Sports, Combat Sports, Sports with Animals,
and generally becomes fine-grained by the leaf level. For
example, our dataset contains 6 different types of bowling,
7 different types of American football and 23 types of bil-
liards.

There are 1000-3000 videos per class and approximately
5% of the videos are annotated with more than one class.
The annotations are produced automatically by analyzing
the text metadata surrounding the videos. Thus, our data is
weakly annotated on two levels: first, the label of a video
may be wrong if the tag prediction algorithm fails or if the
provided description does not match the video content, and
second, even when a video is correctly annotated it may still
exhibit significant variation on the frame level. For exam-
ple, a video tagged as soccer may contain several shots of
the scoreboard, interviews, news anchors, the crowd, etc.

We split the dataset by assigning 70% of the videos to
the training set, 10% to a validation set and 20% to a test
set. As YouTube may contain duplicate videos, it is pos-
sible that the same video could appear in both the training
and test set. To get an idea about the extent of this prob-
lem we processed all videos with a near-duplicate finding
algorithm on the frame level and determined that only 1755
videos (out of 1 million) contain a significant fraction of
near-duplicate frames. Furthermore, since we only use a
random collection of up to 100 half-second clips from ev-
ery video and our videos are 5 minutes and 36 seconds in
length on average, it is unlikely that the same frames occur
across data splits.

Training. We trained our models over a period of one
month, with models processing approximately 5 clips per
second for full-frame networks and up to 20 clips per sec-
ond for multiresolution networks on a single model replica.
The rate of 5 clips per second is roughly 20 times slower
than what one could expect from a high-end GPU, but we
expect to reach comparable speeds overall given that we use
10-50 model replicas. We further estimate the size of our
dataset of sampled frames to be on the order of 50 million
examples and that our networks have each seen approxi-
mately 500 million examples throughout the training period
in total.

Video-level predictions. To produce predictions for an
entire video we randomly sample 20 clips and present each
clip individually to the network. Every clip is propagated
through the network 4 times (with different crops and flips)

Figure 4: Predictions on Sports-1M test data. Blue (first row) indicates ground truth label and the bars below show model
predictions sorted in decreasing confidence. Green and red distinguish correct and incorrect predictions, respectively.

Model Clip Hit@1 Video Hit@1 Video Hit@5
Feature Histograms + Neural Net – 55.3 –
Single-Frame 41.1 59.3 77.7
Single-Frame + Multires 42.4 60.0 78.5
Single-Frame Fovea Only 30.0 49.9 72.8
Single-Frame Context Only 38.1 56.0 77.2
Early Fusion 38.9 57.7 76.8
Late Fusion 40.7 59.3 78.7
Slow Fusion 41.9 60.9 80.2
CNN Average (Single+Early+Late+Slow) 41.4 63.9 82.4

Table 1: Results on the 200,000 videos of the Sports-1M test set. Hit@k values indicate the fraction of test samples that
contained at least one of the ground truth labels in the top k predictions.

and the network class predictions are averaged to produce a
more robust estimate of the class probabilities. To produce
video-level predictions we opted for the simplest approach
of averaging individual clip predictions over the durations
of each video. We expect more elaborate techniques to fur-
ther improve performance but consider these to be outside
of the scope of the paper.

Feature histogram baselines. In addition to comparing
CNN architectures among each other, we also report the ac-
curacy of a feature-based approach. Following a standard
bag-of-words pipeline we extract several types of features
at all frames of our videos, discretize them using k-means
vector quantization and accumulate words into histograms
with spatial pyramid encoding and soft quantization. Ev-
ery histogram is normalized to sum to 1 and all histograms
are concatenated into a 25,000 dimensional video-level fea-

ture vector. Our features are similar to Yang & Toderici
[27] and consist of local features (HOG [5], Texton [24],
Cuboids [8], etc.) extracted both densely and at sparse
interest points, as well as global features (such as Hue-
Saturation, Color moments, number of faces detected). As
a classifier we use a multilayer neural network with Rec-
tified Linear Units followed by a Softmax classifier. We
found that a multilayer network performs consistently and
significantly better than linear models on separate validation
experiments. Furthermore, we performed extensive cross-
validations across many of the network’s hyperparameters
by training multiple models and choosing the one with best
performance on a validation set. The tuned hyper parame-
ters include the learning rate, weight decay, the number of
hidden layers (between 1-2), dropout probabilities and the

Figure 5: Examples that illustrate qualitative differences between single-frame network and Slow Fusion (motion-aware)
network in the same color scheme as Figure 4. A few classes are easier to disambiguate with motion information (left three).

Figure 3: Filters learned on first layer of a multiresolution
network. Left: context stream, Right: fovea stream. No-
tably, the fovea stream learns grayscale, high-frequency fea-
tures while the context stream models lower frequencies and
colors. GIFs of moving video features can be found on our
website (linked on first page).

number of nodes in all layers.
Quantitative results. The results for the Sports-1M

dataset test set, which consists of 200,000 videos and
4,000,000 clips, are summarized in Table 1. As can be
seen from the table, our networks consistently and signif-
icantly outperform the feature-based baseline. We empha-
size that the feature-based approach computes visual words
densely over the duration of the video and produces predic-
tions based on the entire video-level feature vector, while
our networks only see 20 randomly sampled clips individ-
ually. Moreover, our networks seem to learn well despite
significant label noise: the training videos are subject to
incorrect annotations and even the correctly-labeled videos
often contain a large amount of artifacts such as text, ef-
fects, cuts, and logos, none of which we attempted to filter
out explicitly.

Compared to the wide gap relative to the feature-based
baseline, the variation among different CNN architectures
turns out to be surprisingly insignificant. Notably, the
single-frame model already displays strong performance.
Furthermore, we observe that the foveated architectures are
between 2-4× faster in practice due to reduced input dimen-
sionality. The precise speedups are in part a function of the
details of model partitioning and our implementation, but in
our experiments we observe a speedup during training of 6
to 21 clips per second (3.5x) for the single-frame model and
5 to 10 clips per second (2x) for the Slow Fusion model.

Contributions of motion. We conduct further exper-

Sports class ∆ AP ∆ AP Sports class
Juggling Club 0.12 -0.07 Short Track Motor Racing
Pole Climbing 0.10 -0.07 Road Racing

Mountain Unicycling 0.08 -0.07 Jeet Kune Do
Tricking 0.07 -0.06 Paintball
Footbag 0.07 -0.06 Freeride

Skipping Rope 0.06 -0.06 Cricket
Rope Climbing 0.06 -0.06 Wrestling

Slacklining 0.05 -0.06 Modern Pentathlon
Tee Ball 0.05 -0.06 Krav Maga

Sheepdog Trial 0.05 -0.05 Rally Cross

Table 2: Classes for which a (motion-aware) Slow Fusion
CNN performs better than the single-frame CNN (left) and
vice versa (right), as measured by difference in per-class
average precision.

iments to understand the differences between the single-
frame network and networks that have access to motion in-
formation. We choose the Slow Fusion network as a rep-
resentative motion-aware network because it performs best.
We compute and compare the per-class average precision
for all Sports classes and highlight the ones that exhibit
largest differences (Table 2). Manually inspecting some of
the associated clips (Figure 5), we qualitatively observe that
the motion-aware network clearly benefits from motion in-
formation in some cases, but these seem to be relatively un-
common. On the other hand, balancing the improvements
from access to motion information, we observe that motion-
aware networks are more likely to underperform when there
is camera motion present. We hypothesize that the CNNs
struggle to learn complete invariance across all possible an-
gles and speeds of camera translation and zoom.

Qualitative analysis. Our learned features for the first
convolutional layer can be inspected on Figure 3. In-
terestingly, the context stream learns more color features
while the high-resolution fovea stream learns high fre-
quency grayscale filters.

As can be seen on Figure 4, our networks produce in-
terpretable predictions and generally make reasonable mis-
takes. Further analysis of the confusion matrix (attached
in the supplementary material) reveals that most errors are
among the fine-grained classes of our dataset. For exam-
ple, the top 5 most commonly confused pairs of classes are
deer hunting vs. hunting, hiking vs. backpacking, powered
paragliding vs. paragliding, sledding vs. toboggan, and bu-
jinkan vs. ninjutsu.

Model 3-fold Accuracy
Soomro et al [22] 43.9%
Feature Histograms + Neural Net 59.0%
Train from scratch 41.3%
Fine-tune top layer 64.1%
Fine-tune top 3 layers 65.4%
Fine-tune all layers 62.2%

Table 3: Results on UCF-101 for various Transfer Learning
approaches using the Slow Fusion network.

4.2. Transfer Learning Experiments on UCF-101
The results of our analysis on the Sports-1M dataset in-

dicate that the networks learn powerful motion features. A
natural question that arises is whether these features also
generalize to other datasets and class categories. We ex-
amine this question in detail by performing transfer learn-
ing experiments on the UCF-101 [22] Activity Recognition
dataset. The dataset consists of 13,320 videos belonging
to 101 categories that are separated into 5 broad groups:
Human-Object interaction (Applying eye makeup, brush-
ing teeth, hammering, etc.), Body-Motion (Baby crawling,
push ups, blowing candles, etc.), Human-Human interac-
tion (Head massage, salsa spin, haircut, etc.), Playing In-
struments (flute, guitar, piano, etc.) and Sports. This group-
ing allows us to separately study the performance improve-
ments on Sports classes relative to classes from unrelated
videos that are less numerous in our training data.

Transfer learning. Since we expect that CNNs learn
more generic features on the bottom of the network (such
as edges, local shapes) and more intricate, dataset-specific
features near the top of the network, we consider the fol-
lowing scenarios for our transfer learning experiments:

Fine-tune top layer. We treat the CNN as a fixed feature
extractor and train a classifier on the last 4096-dimensional
layer, with dropout regularization. We found that as little as
10% chance of keeping each unit active to be effective.

Fine-tune top 3 layers. Instead of only retraining the fi-
nal classifier layer, we consider also retraining both fully
connected layers. We initialize with a fully trained Sports
CNN and then begin training the top 3 layers. We intro-
duce dropout before all trained layers, with as little as 10%
chance of keeping units active.

Fine-tune all layers. In this scenario we retrain all net-
work parameters, including all convolutional layers on the
bottom of the network.

Train from scratch. As a baseline we train the full net-
work from scratch on UCF-101 alone.

Results. To prepare UCF-101 data for classification we
sampled 50 clips from every video and followed the same
evaluation protocol as for Sports across the 3 suggested
folds. We reached out to the authors of [22] to obtain the
YouTube video IDs of UCF-101 videos, but unfortunately

Group mAP
from

scratch

mAP
fine-tune

top 3

mAP
fine-tune

top
Human-Object Interaction 0.26 0.55 0.52
Body-Motion Only 0.32 0.57 0.52
Human-Human Interaction 0.40 0.68 0.65
Playing Musical Instruments 0.42 0.65 0.46
Sports 0.57 0.79 0.80
All groups 0.44 0.68 0.66

Table 4: Mean Average Precision of the Slow Fusion net-
work on UCF-101 classes broken down by category groups.

these were not available and hence we cannot guarantee that
the Sports-1M dataset has no overlap with UCF-101. How-
ever, these concerns are somewhat mitigated as we only use
a few sampled clips from every video.

We use the Slow Fusion network in our UCF-101 exper-
iments as it provides the best performance on Sports-1M.
The results of the experiments can be seen on Table 3. In-
terestingly, retraining the softmax layer alone does not per-
form best (possibly because the high-level features are too
specific to sports) and the other extreme of fine-tuning all
layers is also not adequate (likely due to overfitting). In-
stead, the best performance is obtained by taking a balanced
approach and retraining the top few layers of the network.
Lastly, training the entire network from scratch consistently
leads to massive overfitting and dismal performance.

Performance by group. We further break down our per-
formance by 5 broad groups of classes present in the UCF-
101 dataset. We compute the average precision of every
class and then compute the mean average precision over
classes in each group. As can be seen from Table 4, large
fractions of our performance can be attributed to the Sports
categories in UCF-101, but the other groups still display im-
pressive performance considering that the only way to ob-
serve these types of frames in the training data is due to label
noise. Moreover, the gain in performance when retraining
only the top to retraining the top 3 layers is almost entirely
due to improvements on non-Sports categories: Sports per-
formance only decreases from 0.80 to 0.79, while mAP im-
proves on all other categories.

5. Conclusions
We studied the performance of convolutional neural net-

works in large-scale video classification. We found that
CNN architectures are capable of learning powerful fea-
tures from weakly-labeled data that far surpass feature-
based methods in performance and that these benefits are
surprisingly robust to details of the connectivity of the ar-
chitectures in time. Qualitative examination of network out-
puts and confusion matrices reveals interpretable errors.

Our results indicate that while the performance is not
particularly sensitive to the architectural details of the con-
nectivity in time, a Slow Fusion model consistently per-
forms better than the early and late fusion alternatives. Sur-

prisingly, we find that a single-frame model already dis-
plays very strong performance, suggesting that local motion
cues may not be critically important, even for a dynamic
dataset such as Sports. An alternative theory is that more
careful treatment of camera motion may be necessary (for
example by extracting features in the local coordinate sys-
tem of a tracked point, as seen in [25]), but this requires
significant changes to a CNN architecture that we leave for
future work. We also identified mixed-resolution architec-
tures that consist of a low-resolution context and a high-
resolution fovea stream as an effective way of speeding up
CNNs without sacrificing accuracy.

Our transfer learning experiments on UCF-101 suggest
that the learned features are generic and generalize other
video classification tasks. In particular, we achieved the
highest transfer learning performance by retraining the top
3 layers of the network.

In future work we hope to incorporate broader categories
in the dataset to obtain more powerful and generic fea-
tures, investigate approaches that explicitly reason about
camera motion, and explore recurrent neural networks as
a more powerful technique for combining clip-level predic-
tions into global video-level predictions.

Acknowledgments: We thank Saurabh Singh, Abhinav
Shrivastava, Jay Yagnik, Alex Krizhevsky, Quoc Le, Jeff
Dean and Rajat Monga for helpful discussions.

References
[1] M. Baccouche, F. Mamalet, C. Wolf, C. Garcia, and

A. Baskurt. Sequential deep learning for human action
recognition. In Human Behavior Understanding, pages 29–
39. Springer, 2011. 2, 3

[2] D. Ciresan, A. Giusti, J. Schmidhuber, et al. Deep neural net-
works segment neuronal membranes in electron microscopy
images. In NIPS, 2012. 1

[3] L. N. Clement Farabet, Camille Couprie and Y. LeCun.
Learning hierarchical features for scene labeling. PAMI,
35(8), 2013. 1, 2

[4] C. Couprie, C. Farabet, L. Najman, and Y. LeCun. Indoor
semantic segmentation using depth information. Internatinal
Conference on Learning Representation, 2013. 2

[5] N. Dalal and B. Triggs. Histograms of oriented gradients for
human detection. In CVPR, volume 1, 2005. 5

[6] J. Dean, G. Corrado, R. Monga, K. Chen, M. Devin, Q. V.
Le, M. Z. Mao, M. Ranzato, A. Senior, P. Tucker, K. Yang,
and A. Y. Ng. Large scale distributed deep networks. In
NIPS, 2012. 4

[7] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-
Fei. Imagenet: A large-scale hierarchical image database. In
CVPR, 2009. 2

[8] P. Dollár, V. Rabaud, G. Cottrell, and S. Belongie. Behav-
ior recognition via sparse spatio-temporal features. In Inter-
national Workshop on Visual Surveillance and Performance
Evaluation of Tracking and Surveillance, 2005. 2, 5

[9] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich fea-
ture hierarchies for accurate object detection and semantic
segmentation. In CVPR, 2014. 1, 2

[10] S. Ji, W. Xu, M. Yang, and K. Yu. 3D convolutional neural
networks for human action recognition. PAMI, 35(1):221–
231, 2013. 2, 3

[11] A. Krizhevsky, I. Sutskever, and G. Hinton. Imagenet clas-
sification with deep convolutional neural networks. In NIPS,
2012. 1, 2, 3, 4

[12] I. Laptev. On space-time interest points. IJCV, 64(2-3):107–
123, 2005. 2

[13] I. Laptev, M. Marszalek, C. Schmid, and B. Rozenfeld.
Learning realistic human actions from movies. In CVPR,
2008. 2

[14] Q. V. Le, W. Y. Zou, S. Y. Yeung, and A. Y. Ng. Learn-
ing hierarchical invariant spatio-temporal features for action
recognition with independent subspace analysis. In CVPR,
2011. 2

[15] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-
based learning applied to document recognition. Proceed-
ings of the IEEE, 86(11):2278–2324, 1998. 1, 2

[16] J. Liu, J. Luo, and M. Shah. Recognizing realistic actions
from videos “in the wild”. In CVPR, 2009. 2

[17] J. C. Niebles, C.-W. Chen, and L. Fei-Fei. Modeling tempo-
ral structure of decomposable motion segments for activity
classification. In ECCV, pages 392–405. Springer, 2010. 2

[18] A. S. Razavian, H. Azizpour, J. Sullivan, and S. Carls-
son. CNN features off-the-shelf: an astounding baseline for
recognition. arXiv preprint arXiv:1403.6382, 2014. 1, 2

[19] P. Sermanet, S. Chintala, and Y. LeCun. Convolutional neu-
ral networks applied to house numbers digit classification. In
ICPR, 2012. 2

[20] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus,
and Y. LeCun. OverFeat: Integrated recognition, localization
and detection using convolutional networks. arXiv preprint
arXiv:1312.6229, 2013. 1, 2

[21] J. Sivic and A. Zisserman. Video Google: A text retrieval
approach to object matching in videos. In ICCV, 2003. 2

[22] K. Soomro, A. R. Zamir, and M. Shah. UCF101: A dataset
of 101 human actions classes from videos in the wild. arXiv
preprint arXiv:1212.0402, 2012. 2, 7

[23] G. W. Taylor, R. Fergus, Y. LeCun, and C. Bregler. Con-
volutional learning of spatio-temporal features. In ECCV.
Springer, 2010. 2

[24] M. Varma and A. Zisserman. A statistical approach to tex-
ture classification from single images. IJCV, 62(1-2):61–81,
2005. 5

[25] H. Wang, A. Klaser, C. Schmid, and C.-L. Liu. Action recog-
nition by dense trajectories. In CVPR. IEEE, 2011. 2, 8

[26] H. Wang, M. M. Ullah, A. Klaser, I. Laptev, C. Schmid, et al.
Evaluation of local spatio-temporal features for action recog-
nition. In BMVC, 2009. 2

[27] W. Yang and G. Toderici. Discriminative tag learning on
youtube videos with latent sub-tags. In CVPR, 2011. 5

[28] M. D. Zeiler and R. Fergus. Visualizing and under-
standing convolutional neural networks. arXiv preprint
arXiv:1311.2901, 2013. 1, 3