tusen-ai / simpledet

A Simple and Versatile Framework for Object Detection and Instance Recognition
Apache License 2.0
3.09k stars 488 forks source link

Excuse me, is the trident net's anchor setting too sparse? #288

Closed YilanWang closed 4 years ago

YilanWang commented 4 years ago

I read the source code of tridentnet, and got some confusion. If stride is 16, then the smallest scale anchor (scale = 2) will be located on the feature map with a very sparse distribution. When the area of retinanet or fpn is the smallest (stride=4, scale=8, length=32, regardless of ratio shift), their distribution on P3 is very dense. However, the APs of tridentnet is greater than FPN. How does such an anchor setting affect the results? Many thanks for your answers!

RogerChern commented 4 years ago

There are many underlying detector design tradeoffs related. Too be short, adding more small anchors is not very rewarding due to the limited resolution of the raw images in COCO.

YilanWang commented 4 years ago

There are many underlying detector design tradeoffs related. Too be short, adding more small anchors is not very rewarding due to the limited resolution of the raw images in COCO.

Many thanks for your reply! Did you try trident-block in single-stage framework, Can it get good results?

RogerChern commented 4 years ago

Some fellow researchers have tried it out on SSD and it worked.

On Thu, Jan 9, 2020 at 2:21 PM wang yilan notifications@github.com wrote:

There are many underlying detector design tradeoffs related. Too be short, adding more small anchors is not very rewarding due to the limited resolution of the raw images in COCO.

Many thanks for your reply! Did you try trident-block in single-stage framework, Can it get good results?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/TuSimple/simpledet/issues/288?email_source=notifications&email_token=ABGODH3JXH2VZBAIGJJDLA3Q42655A5CNFSM4KB4KIMKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEIPEJNI#issuecomment-572409013, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABGODH276ZHBIZ3NAF2HFWTQ42655ANCNFSM4KB4KIMA .

YilanWang commented 4 years ago

Cooool!