Closed rahman-mdatiqur closed 4 years ago
Hi, I refer this post processing tip/trick from another action detection code repository (but sorry I forgot which one, I should have written a comment... it should be a public code repo with paper published in recent 2~3 years). I didn't highlight this operation because as I commented here https://github.com/HYPJUDY/Decouple-SSAD/blob/9a78cea6be27f9cc7cd512e027f00526be5ecad8/operations.py#L410 It didn't improve the performance a lot, especially when doing this operation more than twice, and it cost more time. You can try to remove this operation and compare the performance.
Hello @HYPJUDY thanks for the reply. Yes, you are right in saying that performance does not vary much. I just used the top-1 predictions and the difference is about 0.7 mAP.
Hello @HYPJUDY ,
For post-processing the detections, you seem to be using the top-3 predictions for each class in the function post_process, whereas, in the activity detection literature including the original SSAD paper, only the top prediction is used.
Can you please provide some insights about choosing the top-3 predictions? Do you also do the same for ActivityNet dataset?
I would highly appreciate your response. Thanks.