pyxploiter / Video-Classification-UCF101

Video classification using Resnet (2+1)D on UCF101 Dataset

Installations

• Python 3.8
  • numpy==1.19.2
  • tensorboard==2.4.0
  • torch==1.7.0+cu92
  • torchvision==0.8.1+cu92

or you can install above packages using pip:

pip install -r requirements

Data

UCF101 Action Recognition dataset is used.

Click here to download the dataset

Click here to download the labels

Pretrained Model

Click here to download the pretrained model

Model is trained on following 15 classes from UCF101: ['ApplyEyeMakeup', 'BabyCrawling', 'CleanAndJerk', 'Diving', 'Fencing', 'GolfSwing', 'Haircut', 'IceDancing', 'JavelinThrow', 'Kayaking', 'LongJump', 'MilitaryParade', 'Nunchucks', 'ParallelBars', 'Rafting']

Prepare Data for Training

This script reads the names of classes from "<LABEL_DIR>/classInd.txt" file and copies the videos from <DATA_DIR> to corresponding class folders in the new directory named filtered_data.

usage: prepare_data.py [-h] -d DATA_DIR -l LABEL_DIR

optional arguments:
  -h, --help            show this help message and exit
  -d DATA_DIR, --data_dir DATA_DIR
                        Path to data videos directory
  -l LABEL_DIR, --label_dir LABEL_DIR
                        path to action recognition labels directory

Sample command:

python prepare_data.py \
	--data_dir UCF-101
	--label_dir ucfTrainTestlist

Training

This script trains the model on the data in <DATA_DIR> using the train test splits provided in <LABEL_DIR>

usage: train.py [-h] [-d DATA_DIR] [-l LABEL_DIR] [-o OUTPUT_DIR]
                [-ne NUM_EPOCHS] [-bs BATCH_SIZE] [-fc FRAMES_PER_CLIP]
                [-sc STEP_BETWEEN_CLIPS]

optional arguments:
  -h, --help            show this help message and exit
  -d DATA_DIR, --data DATA_DIR
                        Path to training data videos
  -l LABEL_DIR, --labels LABEL_DIR
                        path to action recognition labels
  -o OUTPUT_DIR, --output OUTPUT_DIR
                        Output path for saving model.
  -ne NUM_EPOCHS, --num_epochs NUM_EPOCHS
                        Number of epochs.
  -bs BATCH_SIZE, --batch_size BATCH_SIZE
                        Batch size of data
  -fc FRAMES_PER_CLIP, --frames_per_clip FRAMES_PER_CLIP
                        framers per video clip
  -sc STEP_BETWEEN_CLIPS, --step_between_clips STEP_BETWEEN_CLIPS
                        steps between video clips

Sample command:

python train.py \
	--data filtered_data \
	--labels ucfTrainTestlist \
	--output saved_model \
	--num_epochs 1 \
	--batch_size 8 \
	--frames_per_clip 5 \
	--step_between_clips 1

Testing

This script tests the model on the data in <DATA_DIR> using the test splits provided in <LABEL_DIR>

usage: test.py [-h] [-d DATA_DIR] [-l LABEL_DIR] [-i SAVED_MODEL]
                [-bs BATCH_SIZE] [-fc FRAMES_PER_CLIP]
                [-sc STEP_BETWEEN_CLIPS]

optional arguments:
  -h, --help            show this help message and exit
  -d DATA_DIR, --data DATA_DIR
                        path to training data videos
  -l LABEL_DIR, --labels LABEL_DIR
                        path to action recognition labels
  -m SAVED_MODEL, --saved_model SAVED_MODEL
                        path to saved model.
  -bs BATCH_SIZE, --batch_size BATCH_SIZE
                        Batch size of data
  -fc FRAMES_PER_CLIP, --frames_per_clip FRAMES_PER_CLIP
                        framers per video clip
  -sc STEP_BETWEEN_CLIPS, --step_between_clips STEP_BETWEEN_CLIPS
                        steps between video clips

Sample command:

python test.py \
	--data data \
	--labels ucfTrainTestlist \
	--saved_model saved_model/best_model.pth \
	--batch_size 16 \
	--frames_per_clip 5 \
	--step_between_clips 1

Infer

This script loads the trained video classification model from <SAVED_MODEL> path and shows the prediction on the video file from <VIDEO_FILE> path.

usage: infer.py [-h] [-v VIDEO_FILE] [-m SAVED_MODEL] [-l LABEL_DIR]
                [-fc FRAMES_PER_CLIP] [-sc STEP_BETWEEN_CLIPS]

optional arguments:
  -h, --help            show this help message and exit
  -v VIDEO_FILE, --video_file VIDEO_FILE
                        path to video file
  -m SAVED_MODEL, --saved_model SAVED_MODEL
                        path to saved model.
  -l LABEL_DIR, --label_dir LABEL_DIR
                        path to action recognition labels directory
  -fc FRAMES_PER_CLIP, --frames_per_clip FRAMES_PER_CLIP
                        framers per video clip
  -sc STEP_BETWEEN_CLIPS, --step_between_clips STEP_BETWEEN_CLIPS
                        steps between video clips

Sample command:

python infer.py \
	--video_file test.avi
	--saved_model saved_model/best_model.pth
	--label_dir ucfTrainTestlist