search

MIV-XJTU / ARTrack

Apache License 2.0
238 stars 34 forks source link

Inference slow #85

Closed Ocean23333 closed 1 month ago

Ocean23333 commented 1 month ago

Hello, i followed the Readme.md and only made a slight modification(pycocotools==2.0.8 ). When i ran the test.py on 4090 gpus with the script： python test.py artrack_seq artrack_seq_256_full --dataset lasot --threads 0 --num_gpus 1 or python test.py artrack_seq artrack_seq_256_full --dataset lasot --threads 8 --num_gpus 2 i got about 14 fps。 I dont know where‘s problem up. I hope you can provide some suggestions, thank you in advance!

AlexDotHam commented 1 month ago

The question about the inference speed may come from the CPU, i use the official code in 3090 server, i get about 18 FPS, when use the inference test code (like in OSTrack), i get the FPS about 26.

AlexDotHam commented 1 month ago

By the way, the inference speed of ARTrack is really slow, it comes from the inter-frame autoregressive, in other words you need to decode the trajectory sequence from x_min, to y_max. The four time decode make the speed unacceptable.

Ocean23333 commented 1 month ago

Thank you for your rapid reply, i am trying to run the inference in the V2 version(temporal code). Looking forward to the release of your V3 paper and code.Great work!