zhyj3038 / ActivityNet_task_1

data_preparation+TSN+BSN

ActivityNet_task_1

This repository holds the codes for ActivityNet Challenge Task 1: Temporal Action Proposals

Related repositories:

Temporal Segment Networks
TSN PyTorch
BSN-boundary-sensitive-network.pytorch
anet2016 CUHK

*Update

DPN92 RGB+Flow nonrescale feature Baidu_pan_passwd_x8ff

Usage Guide

Setting

conda env create -f environment.yml

Data Preparation

Extract Frames and Optical Flow Images

TSN

Training TSN

RGB&Flow

cd ./TSN && sh run.sh

Extracting video's feature

RGB&Flow

sh extract_feature.sh

BSN

Rescaling RGB and Flow's features and concatenate them

cd ./BSN
python data_process.py arch ../TSN/RGB_feature_folder ../TSN/Flow_feature_folder num_works

Training, evaluating and testing BSN

sh bsn_3fold_diff_backbone.sh

Final result proposal

./BSN/output/result_proposal.json

Ensemble

Soft-NMS layer ensemble

Naive greed search to find a combination.

   mv ./BSN/ensemble/* ./BSN
   # need to add

Demo

Put TSN RGB trained model in the folder ./demo_activity
Download: dpn92_RGB_k600_fold_2_rgb_model_best.pth.tar, google_drive or baidu_pan_passwd_38cw

cd demo_activity && sh run.sh VIDEO_PATH