The performance of SlowFastR101+8×8+K600 is lower than official data.
rogercmq opened this issue · comments
rogercmq commented
In the official SlowFast repo, the performance of SlowFastR101+8×8+K600 on AVA2.2 is 29.1 mAP (without E2E) . See https://github.com/facebookresearch/SlowFast/blob/main/MODEL_ZOO.md#ava
However, in your paper, the baseline SlowFastR101+8×8+K600 mAP is 27.4. Is it your re-implementation result?
Shoufa Chen commented
A similar issue can be found here: facebookresearch/SlowFast#141.
In summary, the official repo does not release the training schedule for R101. We train R101 by ourselves.
A similar performance is also achieved here.
rogercmq commented
Got it! Thx!