SSA2D

Code for Single Shot Actor-Action Detection in Videos (SSA2D). Paper accessible at [WACV 2021 Proceedings] (https://openaccess.thecvf.com/content/WACV2021/papers/Rana_We_Dont_Need_Thousand_Proposals_Single_Shot_Actor-Action_Detection_in_WACV_2021_paper.pdf)

Description

This is an implementation of SSA2D on A2D and VidOR datasets. It is built using the Keras library. The datasets have to be downloaded separately.

Setup

Download the dataset and assign their paths in respective dataloader. For VidOR, assign data and annotation path in dataloader_vidor.py file. VidOR uses some processed annotation files, available in data folder.

Training

Run train_ssa2d_vidor.py to train SSA2D on VidOR. The trained models and log files will be saved in specified folders.

To-Do

1. Add clean code for A2D dataloader.
2. Add clean code for full evaluation.
3. Generalized dataset format for both datasets.