search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.8k stars 7.96k forks source link

Yolo spp vs 5 yolo layers #2714

Open VisionEp1 opened 5 years ago

VisionEp1 commented 5 years ago

Hi,

i read that yolo spp is better then yolo "normal" in any cases.

When should we use:

Full-model: 5 yolo layers: https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov3_5l.cfg

vs

Spatial-full-model: 3 yolo layers: https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov3-spp.cfg

Is there also a 5 yolo layers spp model ?

thanks

kmsravindra commented 5 years ago

i read that yolo spp is better then yolo "normal" in any cases.

I am not sure if this is always true. In my case yolov3 gave superior results compared to yolov3-spp model with all other parameters remaining the same. I think spp model is more suited if the nature of objects are more distinguishable from their local features..

A possible reason I think I had this anomaly is because I had similar looking objects (but different ones) in many of my images...Maybe in such cases, global feature extraction might be better than local feature extraction, since local features of look alike objects match with annotated actual objects, thus resulting in triggering more false negatives on actual objects...

If anyone had similar experience, then please share...

duongdqq commented 4 years ago

i read that yolo spp is better then yolo "normal" in any cases.

I am not sure if this is always true. In my case yolov3 gave superior results compared to yolov3-spp model with all other parameters remaining the same. I think spp model is more suited if the nature of objects are more distinguishable from their local features..

A possible reason I think I had this anomaly is because I had similar looking objects (but different ones) in many of my images...Maybe in such cases, global feature extraction might be better than local feature extraction, since local features of look alike objects match with annotated actual objects, thus resulting in triggering more false negatives on actual objects...

If anyone had similar experience, then please share...

Hi @kmsravindra I think I am having a same problem. I am using yolov3.cfg. I trained my model on a dataset including a significant number of objects. These objects have quite same ratios and have not so much feature to distinguish. And here is my result. I just work with binary classification mAP 81% precision 0.81 recall 0.77 False positive rate 10% Missing False Negative rate = 1 - recall = 0.23 I think I am in your case now. I really want to raise Recall ( that means decrease Missing FNR) I think I have to decrease False negative results. But I do not find a key Do you solve your problem? Can you share your experience