/Improving-Action-Quality-Assessment-using-ResNets-and-Weighted-Aggregation

Code for the paper accessible at https://arxiv.org/abs/2102.10555

Primary LanguagePython

This is the PyTorch code for the following paper presented in IbPRIA 2022:

Improving Action Quality Assessment Using Weighted Aggregation

Arxiv Preprint

The code can be used to reproduce the resutls on MTL-AQA dataset.

Pre-trained models: The resnet's used in our code requires pre-trianed models. These mdels can be found in the following repositoreis:

3D Resnet 34,50,101; (2+1)D-Resnet 34,50,101, all 16 frame input processing and pretrianed on Kinetics Action recognition dataset

(2+1)D Resnets 34 processing 8 and 32 frames, pretrained on kinetics and fine tuned on ig-65m dataset

Requiremetns

Dataset

Download the dataset from here. Extract the necessary frames from the videos. The code expects the frames to be in a folder structure such as this:

~/frames
   /video0_freames
     /frame0.jpg
     .
     .
     .
     /frame1000.jpg
   /video1_frames
     /frame0.jpg
     .
     .
   .
   .
   .

Running the code

Execute the following command:

python train_test_resnet.py --initial_epoch 0

Change the value of initial_epoch to resume the execution from some other epoch.