Update: We've won 3rd Place at the competition! Read Here for details on the competition task and our approaches. We look forward to sharing the Technical Report as soon as possible. Competition Link

This repository contains codes and relevant materials for team ServerDown's submission to 2nd AVA Challenge@IEEE MIPR 2024.

Experiment Results as Presented in Technical Report w/ Codes and Submission Files:

Experiment	Public ROC	Private ROC
Per-Frame Considering Last Three Frames code	0.6873	0.7428
VidSwin code	0.6619	0.7118
Ensemble (Average) (CNN-Transformer and Per Frame) code	0.7459	0.7005
End-to-End CNN-Transformer code	0.6905	0.6820
Ensemble (Weighted) (CNN-Transformer and Per Frame) (Selected Submission on Kaggle) code	0.7560	0.6576
Pretrained-CNN + RNN code	0.6280	0.6571

Given the large size of the model weights, we did not upload them to GitHub. You can download the trained weights from Google Drive and update the paths in the inference codes to reproduce the results.