Action Recognition Evaluation Pipeline on BCN Feature

Required Environment

python 3.8.8
Torch 1.7

Video feature of ActivityNet validation set

Download the video feature of ActivityNet validation set from https://pan.baidu.com/s/1CqYLlkA9mMNWSrsVhb4Jjg (pass code: r8vd)
Download the weights folder from the same URL and put it into this repo (./weights/ckpt_epoch_32.pth)
The video feature of each video clip is extracted by BCN model (implemented by Caffe) with frame stride 8
For more details about BCN model, please refer to https://github.com/FuchenUSTC/BCN
Please refer to ./dataset/anet/anet_val_npy.csv for more details about the clip number and YouTube Video ID of each video
All the video number is 4,926 (5K), all the clip number is 2,066,253 (2M)

Configuration

Modify the eva_root_path in ./config/mlp-anet-infer.yml as the path of the validation feature
Modify the clip_stride in ./config/mlp-anet-infer.yml for different number of sampling clips for evaluation
When clip_stride is -1, the number of sampled clips equals to video number (5K)
When clip_stride is 20, the number of sampled clips is 103,312 (100K)
When clip_stride is 2, the number of sampled clips is about 1M

Evaluation

If the environment and configuration have been set, please run

bash run_eval.sh

The Top-1, Top-3 and Top-5 classification accuracies are recorded on the log folder ./output/mlp-anet-infer

Logs

Please refer to the logs ./output/mlp-anet-infer-5K/log.txt (Top-1: 0.9275) and ./output/mlp-anet-infer-100K/log.txt (Top-1: 0.9252) for more details about the performances on different number of clips (5K and 100K)

FuchenUSTC / BCN-VAL