Temporal Segment Networks

tsn model for action recognition on pytorch

Build =

build opencv 2.4.13 dense_flow
```
 $ bash build_all.sh
```
"All tools built. Happy experimenting!" build
Download dataset = UCF-101HMDB51

Extract optical flow = $ bash scripts/extract_optical_flow.sh DATASET_PATH OUT_PATH NUMBER_OF_WORKER
```
 DATASET_PATH : 

 OUT_PATH 

 NUMBER_OF_WORKER 2
```

frame

4.warped Etract warped flow

tools/build_of.py89--flow_type``warped_tv11

parser.add_argument("--flow_type", type=str, default='warped_tvl1', choices=['tvl1', 'warp_tvl1'])

$ bash scripts/extract_optical_flow.sh DATASET_PATH OUT_PATH NUMBER_OF_WORKER

Label
=

ucf101

  $ bash scripts/build_file_list.sh ucf101 FRAME_PATH

hmdb51

  $ bash scripts/build_file_list.sh hmdb51 FRAME_PATH

FRAME_PATHframe

6.(Inception-BN) Training

ucf101
- rgb
- flow
- rgb-diff
- warped flow

hmdb51

rgb

  $ python main.py hmdb51 RGB <hmdb51_rgb_train_list> <hmdb51_rgb_val_list> 

  	--arch BNInception --num_segments 3 

  	--gd 20 --lr 0.001 --lr_steps 30 60 --epochs 80 

  	-b 128 -j 8 --dropout 0.8 

  	--snapshot_pref ucf101_bninception_ 

  	--gpus 0 1