Implement Yolo based on PANet ~+4 AP@[.5, .95]

[AlexeyAB]

Implement Yolo based on PANet: https://arxiv.org/abs/1803.01534v4

1. There is added very short Bottom-up Path Augmentation - so there are added shortcuts (or conctats) to FPN (Feature Pyramid Networks for Object Detection that is used for Yolo v3), so each of 4 final layers has no more than 10 intermediate layers from any Level.

2. Also there is added Adaptive Feature Pooling Structure

   • either just 3 [shortcut] layers, for element-wise adding of 4 final layers
   • or we can implement pooling layer that uses max operation across channels of two input layers

3. May be use both Deconvolutional layer and Fully-connected layer as in PANet, or just Deconvolutional layer and YOLACT (You Only Look At CoefficienTs) YOLACT: Add real-time Instance Segmentation #3048

---

• PANet (Path Aggregation Network for Instance Segmentation): https://arxiv.org/abs/1803.01534v4
• FPN (Feature Pyramid Networks for Object Detection): https://arxiv.org/abs/1612.03144v2

[AlexeyAB]

Some form of PANet implementation: #3114 (comment)

[AlexeyAB]

Example:

cfg-file: https://github.com/AlexeyAB/darknet/files/3253820/yolo_v3_spp_pan_scale.cfg.txt

pre-trained weights-file for training: http://pjreddie.com/media/files/darknet53.conv.74

training command: ./darknet detector train data/self_driving.data yolo_v3_spp_pan_scale.cfg.txt darknet53.conv.74 -map

detection command: ./darknet detector demo data/self_driving.data yolo_v3_spp_pan_scale.cfg.txt backup/yolov3-tiny_occlusion_track_last.weights self_driving_valid.avi -out_filename out_self_driving_valid.avi

result training weights-file: https://drive.google.com/file/d/14557YbnxPepJIUlrTdaphhNmMAVLsOqq/view this dataset: https://drive.google.com/open?id=1QzXSCkl9wqr73GHFLIdJ2IIRMgP1OnXG

---

example of detection: https://drive.google.com/open?id=1EcHfzgypfk57pQlLl379c-CKvSy5l3zg

example of training Loss & mAP chart: