search

hustvl / YOLOP

You Only Look Once for Panopitic Driving Perception.（MIR2022）
MIT License
1.9k stars 411 forks source link

Perform object detection on segmentation area #101

Open danielzhangau opened 2 years ago

danielzhangau commented 2 years ago

I want to improve the accuracy of object detection by yolop. Is there any way for yolop to directly use the segmented road part for the target detection model after completing the segmentation, or only display the result of the detection target within the segmentation range?

novoforce commented 2 years ago

hi @danielzhangau from what i understand is YoloP has 3 decoder head which takes the features from the common feature extractor(encoder) and perform 3 different different tasks(detection, segmentation, segmentation). So for your query directly via YoloP is not possible. Alternatively you can take the output of the 1 decoder segmentation head and perform another detection on top of that or you might have to play around with the architecture of YoloP by adding some residual layer on the segmentation head and perform detection.