TSN

This repository derives from the Temporal Segment Network for Action Recognition.

Paper: Temporal Segment Network

Preparing dataset

The model has been trained on UCF-101 dataset. The dataset can be obtained from UCF-101

Extracting RGB frames

We need to extract RGB frames to train the model. RGB frames can be extracted by:
bash scripts/extract_frames.sh SRC_DIR OUT_DIR

Creating list files

A list file containing path of the video frames, number of frames in a video and label of the video is required to be passed to the model for training.

List file can be created by:
bash scripts/build_list_file.sh ucf101 FRAME_PATH OUT_FILE_PATH

FRAME_PATH: Path to which the frames were extracted.

Training model with BNInception backbone

python main.py ucf101 RGB TRAIN_SPLIT_FILE_PATH VAL_SPLIT_FILE_PATH --arch BNInception --num_segments 3 --gd 20 --lr 0.001 --lr_steps 30 60 --epochs 80 -b 128 -j 8 --dropout 0.8 --snapshot_pref ucf101_bninception_ --gpus 0 1 2 3

Testing model

python test_models.py ucf101 RGB VAL_SPLIT_FILE_PATH ucf101_bninception__rgb_model_best.pth.tar --arch BNInception --save_scores scores_bninception --workers 4 --gpus 0 1 2 3