TEST_README.md

Segmenting Video Sequences with TDNet

1. Prepare Data

Save video frames with their IDs as names. An examplar video is saved in ./data/vid1.

2. Prepare Model:

• Download pretrained models from the [G-cloud].
• Decompress checkpoint.zip and put the folder into this dirctory.

3. Running:

Test with TD2-PSP50:

python test.py --gpu 1 --model td2-psp50

Test with TD4-PSP18:

python test.py --gpu 1 --model td4-psp18

For performance comparison, you can also run with PSPNet-101:

python test.py --gpu 1 --model psp101

4. Latency/Speed

Model	Latency (on Titan Xp)	Per-frame FLOPs
PSPNet101	~360ms/f	~1.2T/f
TD2-PSP50	~180ms/f	~541G/f
TD4-PSP18	~85ms/f	~239G/f