tiny model struggles with long objects (>= 70% of img width)

[holger-prause]

Hello,
I am training a custom dataset based on yolov4-tiny-3l.cfg, and its working pretty fine for most objects and i am happy with the results(compared to yolov7-tiny its waaaay better). However - the model constantly struggles with very long objects (70% of the image), even with custom anchors and padding the image.

This problem seems to happen across all the classes - so it seems to be a conceptional problem related to anchor boxes. Training with full yolov4 does not help.

@AlexeyAB

• Does it have to do something with the perceptive field?
• Is there anything i can do about it?
• Would putting those long objects into a separate class help (they have same same appearance as the small object of the same type- they are just longer)?
• Are there any config entries i could change?
• Would calculating anchors with a bigger or even fewer -num_of_clusters help?

Any ideas or hints into the right direction would help, the rest i can try out myself.
Thank you very much

[holger-prause]

I was able to reproduce this problem by creating a dataset which contains those long object only.
I.E i have an image with 1900 * 1080 and the objects are around 1400 * 10. When training with default anchors or even with custom anchors - none of them are picked up - if you are lucky - a quarter is detected sometimes.

So this is definitivly a problem - at least for me. Reading a bit about this phenomen on yolov5(which does NOT have this problem) - it seems aspect ratio/anchor box related.

I would be very happy for a workaround on this. In the meantime i tried yolov7, yolox, and yolov5. All of them are behaving fine on the same dataset with those long and thin objects.

I am still a bit shocked that very few people have the same problem and its hard to find any material on this.