meituan / YOLOv6

YOLOv6: a single-stage object detection framework dedicated to industrial applications.
GNU General Public License v3.0
5.72k stars 1.04k forks source link

Need help in utilizing 4 feature pyramids to create 3 feature maps #430

Closed sarmientoj24 closed 2 years ago

sarmientoj24 commented 2 years ago

I saw this diagram as the architecture of YOLOv6's neck. It seems to use 3 feature pyramids from the backbone. However, I have 4 feature pyramids from the backbone I am using to test with YOLOv6. I would like to ask help in the implementation of this.

Basically, it will take 4 feature maps -- C2, C3, C4, and C5 but only has to output P3, P4, and P5. image

shensheng272 commented 2 years ago

Current neck has 3 in 3 out. IF you want to 4 in 3 out, you might just ignore C2 and input C3 C4 C5, so neck is still 3 in 3 out. Or you might merge C2 and C3 (with downsample, conv, add or concat)to a new_C3 and input new_C3 C4 C5. IF you want to 4 in 4 out, you need to adjust neck and head.

Chilicyy commented 2 years ago

@sarmientoj24 You can try to add some operations like RepBlocks+1*1conv+upsample to build P2 outputs . Modify the code in ./yolov6/models/reppan.py.