Detecting smaller objects or objects in the distance
jameslahm opened this issue · 8 comments
Due to that YOLOv10 adopts the different training strategy with others, e.g. YOLOv8, it may thus have different favorable confidence threshold to detect objects. Besides, different thresholds will have no impact on the inference latency of YOLOv10 because it does not rely on NMS. Therefore, we suggest that a smaller threshold can be freely set or tuned for YOLOv10 to detect smaller objects or objects in the distance. For example,
wget https://skalskip-yolo-arena.hf.space/file=/tmp/gradio/048ecf37df927d6d7ed5528076f4133147f797d2/people-walking.png
yolo export model=yolov10m.pt format=onnx opset=13 simplify half=True device=0
# Here, we could use a smaller threshold of 0.05 rather than the default threshold of 0.25 to detect smaller objects
# or objects in the distance.
yolo predict model=yolov10m.onnx source=people-walking.png half conf=0.05
Thank you all! ❤️
Thank you! Your assistance has been instrumental in boosting both the F1-score and the mean Average Precision (mAP) of our trained model.
What is the F1-score and mAP of the model after setting threshold as described above?
@jameslahm @Sompote
Also, can we modify NMS to dynamically change threshold so that it uses a lower confidence threshold for smaller objects?
How to lower confidence level during training for fair mAP comparison between YOLOv10 and other models on datasets with smaller bounding boxes?
how to solve the problem of Misidentification by lowering the threshold?
@jameslahm do you plan to implement yolov10-p2 version ?
+1
+1
but i think, the person on the top left is not a small object, and it is obvious instead, why confidence is so low? the model doesn't converge well?