Overlapped objects
ToshiEAB opened this issue · comments
Dear AlexeyAB
I am using your program (https://github.com/AlexeyAB/Yolo_mark) to train a YOLO model. Specifically, after the training, I want to detect several small fish at once on a video even when they somewhat overlap (e.g., two fish making an "X" shape") with each other. What is a good way of training to achieve that goal? Should I use images showing two (or more) fish overlapping each other and put a label on each of the fish? Or should I use images showing only a single fish and put a label? If it is difficult to detect two separate fish at once when they are making an "X" shape, should I consider the X shape as a "negative" sample and then use a tracker (e.g., Kalman filter) to make a prediction of the motion of each fish.
Any advice would be helpful.
Thanks in advance,
ToshiEAB
@ToshiEAB Hi,
Several overlapped objects with the same class_id is the most difficult case in object detection.
Currently there are 3 features in road-map for implementing:
- Repulsion Loss (CVPR 2018) AlexeyAB/darknet#3113
- Soft-IoU layer (CVPR 2019) AlexeyAB/darknet#4701
- Yolact instance segmentation, each instance of object in its own channel: AlexeyAB/darknet#3048
should I consider the X shape as a "negative" sample and then use a tracker (e.g., Kalman filter) to make a prediction of the motion of each fish.
Yes, you can try to do this.
There is no best practice for this case.