search

HYPJUDY / Decouple-SSAD

Decoupling Localization and Classification in Single Shot Temporal Action Detection
https://arxiv.org/abs/1904.07442
MIT License
96 stars 19 forks source link

Matching the default anchors with gt #11

Closed rahman-mdatiqur closed 4 years ago

rahman-mdatiqur commented 4 years ago

Hello @HYPJUDY ,

when you match the default anchors with gt bboxes, you are checking if the overlap is greater than -0.5 or not as in this line. Since, the overlaps can never be negative, what is the implication of the below line ? mask = tf.logical_and(mask, b_match_scores > -0.5)

Thanks.

rahman-mdatiqur commented 4 years ago

Hello @HYPJUDY , can you please respond?

HYPJUDY commented 4 years ago

I am also confused. It's adopted from the original SSAD code. I only add a comment line when using this part. https://github.com/HYPJUDY/Decouple-SSAD/blob/9a78cea6be27f9cc7cd512e027f00526be5ecad8/operations.py#L80 I've consulted the original coder, he said it's refered from a tensorflow code of SSD and maybe the -0.5 is not necessary. Only mask = tf.logical_and(mask, b_match_scores > 0) is enough. Please provide feedback if it doesn't work. Thanks!

rahman-mdatiqur commented 4 years ago

Thanks much for your detailed feedback. I highly appreciate your quick response. I will let you know how it performs without this line.

rahman-mdatiqur commented 4 years ago

Hello @HYPJUDY ,

I did not find any difference in performance with that line commented out.

Thanks for all your help.

HYPJUDY commented 4 years ago

Thanks for testing this!