Action Recognition in Video

This repo will serve as a playground where I investigate different approaches to solving the problem of action recognition in video.

I will mainly use the UCF-101 dataset.

Setup

$ cd data/              
$ bash download_ucf101.sh     # Downloads the UCF-101 dataset (~7.2 GB)
$ unrar x UCF101.rar          # Unrars dataset
$ unzip ucfTrainTestlist.zip  # Unzip train / test split
$ python3 extract_frames.py   # Extracts frames from the video (~26.2 GB, go grab a coffee for this)

ConvLSTM

The only approach investigated so far. Enables action recognition in video by a bi-directional LSTM operating on frame embeddings extracted by a pre-trained ResNet-152 (ImageNet).

The model is composed of:

• A convolutional feature extractor (ResNet-152) which provides a latent representation of video frames
• A bi-directional LSTM classifier which based on the latent representation of the video predicts the activity depicted

I have made a trained model available here.

Train

$ python3 train.py  --dataset_path data/UCF-101-frames/ \
                    --split_path data/ucfTrainTestlist \
                    --num_epochs 200 \
                    --sequence_length 40 \
                    --img_dim 112 \
                    --latent_dim 512

Test on Video

$ python3 test_on_video.py  --video_path data/UCF-101/SoccerPenalty/v_SoccerPenalty_g01_c01.avi \
                            --checkpoint_model model_checkpoints/ConvLSTM_150.pth

Results

The model reaches a classification accuracy of 91.27% accuracy on a randomly sampled test set, composed of 20% of the total amount of video sequences from UCF-101. Will re-train this model on the offical train / test splits and post results as soon as I have time.