Closed Gooogr closed 4 years ago
After some tests, I understood those results are normal for tiny YOLO architecture. You can improve it if you will use custom train anchors and play around with the detection threshold. In this case, use AlexeyAB`s realization, not pjreddie's one.
Hello! I got very bad detection result for tiny YOLO model, it feels like model is highly underfitted. I used
detection.py
file and changed strings 14 - 16 to use tiny version like:The weights of models were downloaded by links from READ.ME file and were converted to tensorflow format without any errors.
Examples with YOLOv3 and Tiny YOLO v3: