Difference between yolov3-tiny.cfg and yolov3-tiny_3l

[pushkalkatara]

Hi,
How to understand the difference between
yolov3-tiny.cfg and yolov3-tiny_3l

I notice addition from here route, cnn, upsample, route, cnn, cnn, yolo

[AlexeyAB]

@pushkalkatara Hi,

yolov3-tiny_3l uses 3 [yolo]-layers at different scales, while
yolov3-tiny uses 2 [yolo]-layers at different scales

[pushkalkatara]

Can you please share the intuition behind adding an extra yolo layer with different scales. Also, how are the accuracy and fps affected in both models?

[AlexeyAB]

There was idea, that small objects require small features which are closer to input,
while big objects require big feature which are far from input.

For example, [yolo] layer with mask= that reference to smaller anchors= is used for small objects, and we though that it requires shortcut to first layers to get small features.

But now we know, that all [yolo] layers (for small and big objects) require all features (small and big), from this article for PANet: #3175 and paper https://arxiv.org/abs/1803.01534v4

4 colored lines are 4 detection ([yolo] or other) layers:

So: #3114 (comment)

• yolov3-tiny.cfg uses 2 [yolo]-layers - Inference = 5.0 ms | mAP = 32.3%

• yolov3-tiny_3l.cfg.cfg uses 3 [yolo]-layers - Inference = 5.6 ms | mAP = 46.8%

• yolo_v3_tiny_pan.cfg uses 3 [yolo]-layers - Inference = 8.7 ms | mAP = 49.7%

• yolo_v3_spp.cfg uses 3 [yolo]-layers - Inference = 23.5 ms | mAP = 51.8%

• yolo_v3_spp_pan.cfg uses 3 [yolo]-layers - Inference = 33.8 ms | mAP = 58.5%

Read more: #3114 (comment)

[pushkalkatara]

Great. Thanks for the detailed answer. 👍