Action Recognition in Videos Using LRCN

Project overview:

I implemented the LRCN model from the paper [1] and studied the model accuracy on variety of cases. The model was trained on the UCF-101 dataset that holds ~13K movies of 101 different human actions [2]. The following project consist of:

1. Implementation of the LRCN model in PyTorch 1.3.4
2. Training of the model on a sub-set of 55 classes (on Google Cloud Platform).
3. Evaluation of the model using custom tests.
4. A pipeline that labels an arbitrary video for human actions with single frame resolution.

tl;dr

1. The LRCN model can classify the human action occurring in the test set videos.
2. The model classification accuracy decreases during the video fragment where the action type changes. In this case the level of accuracy depends on the number of frames from each class.
3. The LRCN model can generalize on arbitrary videos (e.g. random YouTube videos).

Table of content

• LRCN model overview
• Data Pre-processing
• Test the accuracy of the model on test dataset
• Test the accuracy of the model on a video that contains multiple human actions
• Predicting human action in arbitrary videos

LRCN model overview

The LCRN model enables human action recognition in videos. The video frames are embedded by a pre-trained ResNet-152 (ImageNet) and processed by a bi-directional LSTM layer that outputs the labels of the preformed action. This model implements the CNN-RNN framework.

The input of the network is a short fragment of a video presenting human activity (N frames without the audio). The output is a probability vector of possible labels.

Data Pre-processing

Each video is downsampled and cropped to a single fragment of constant length (number of frames/fragment is a hyper-parameter).

Elaboration on the preprocessing pipeline:

1. Read each video in the raw video dataset using CV2.
2. Downsample the video to 2.5 FPS (adjustable by args.sampling_rate)
3. Extract N consecutive frames from a random location in the video. Discard the rest.

Basic test mode:

Testing the accuracy of the model on the test dataset (comes with the UCF-101 Dataset). The model reached a classification accuracy of 90.5% when tested on the UCF-101 test dataset (Random prediction baseline = 1/55 = 1.8%)

I analyzed the performance of the model by calculating the confusion matrix. The Confusion matrix suggests that the model confuses similar classes such as 'Military Parade' with 'Band marching' and 'Haircut' with 'Blow Dry Hair'.

Could this confusion be an artifact of an unbalanced data, or is it due to class similarity?

First, I examined how many data points are there in each class. As shown in the figure the classes are slightly imbalanced (See ‘Blowing Candles’ class)

To test if the classification accuracy drops with mis-represented class, I plotted the accuracy of each class and marked in red all the misrepresented ones. A class is defined as misrepresented if its volume is smaller by 10% than the average.
I don’t see significant correlation due to contradicting examples such as ‘Jumping rope’ and ‘Apply Lipstick’.

Diverse human actions video test mode:

This test checks the model performance on videos that contain change of a human action (for example the first part describes one action that changes to another).

I trained the model to predict human action based on fragments of 5 frames. Can the model predict the correct action when 2 of the frames describes 'applying makeup' and the other 3 describes 'crawling baby'?

This test was applied to 16 concatenated video fragments (5 frames each, total test length 80 frames) The fragments were randomly chosen from the test dataset. The model predicts the label of each consecutive 5 frames (‘sliding-window’ of size 5). The true label of each sliding window is set according to the maximal frame count from a single action: Each fragment can span up to 2 different actions, see image below

In this test the model reached a classification accuracy of 73.6%. The inferior accuracy is explained by the frequent action transitions.

* Correct classification is marked in green. False classification is marked in another color (each class has its own color).

I analyzed the accuracy of the model as a function of the number of frames from the first action in the 5 frame window.

The results indicate that the classification accuracy depend on the number of frames from each class. When all frames belong to one action (N=5 frames) the model has high classification accuracy. The accuracy decreases with the number of frames of one class, down to ~60% in case of almost equal frame count (in the case of 3 frames from the first action and 2 frames from the second action)

What does the model predict when it misclassifies an action? To check this, I plotted the predicted label VS. the true label, colored as follows:

1. Light orange: The true and the predicted labels are identical.
2. Red: The predict label is not included in the 2 actions occurring in the video.
3. Light-blue: the true label is one of the 2 actions occurring in the video.
4. Royal-blue: the predicted label is one of the 2 actions occurring in the video

   

The above results show that in 40% of the cases (6/15) the model predicts the second action in the window although the majority of the frames are from the first action. This indicates that the last frames in the window could have higher influence in the classification of the action. This observation can be studied by:

1. Repeating this experiment multiple times with different videos and see if the results are consistent.
2. Investigating how the output of the LSTM changes with the sliding window.
3. Investigating the weight of the LSTM forget gate.
4. Checking how bidirectional LSTM or attention influence the model performance.

Predicting human action in arbitrary videos

The performance of the model is tested on videos taken from YouTube.

Similarly to the previous test, the video is down-sampled to match the FPS of the training data set. The model produces a sequence of labels by predicting the action accruing in each N-frames of the sliding window.

The model classified correctly 91.8% of the video frames.

Reference

1. Donahue, Jeffrey, et al. "Long-term recurrent convolutional networks for visual recognition and description." Proceedings of the IEEE conference on computer vision and pattern recognition. 2015.
2. Soomro, Khurram, Amir Roshan Zamir, and Mubarak Shah. "UCF101: A dataset of 101 human actions classes from videos in the wild." arXiv preprint arXiv:1212.0402 (2012).