Jumabek
custom anchors not improving the performance
I use the provided scripts to generate the anchors on the wider dataset. However, the performance is not improving much compared to the original anchors. Have you observed the similar trend? Thanks.
I also started training wider yesterday. My purpose was to see how much improvement can be gained with custom anchors. Give me 2 more days. I will update you on that
Sure. Waiting for your feedback. I evaluated the detector on FDDB as well as on individual samples. Not as good as the original anchors. I have also tried to increase the anchor number to 9, no improvement is observed.
Hi @fishman2008 , sorry for getting back late.
I stuck in my evaluations, cuz valid wider keeps giving error while evaluating. That's because of new function print_wider_detections that I wrote, but I have no idea how to fix it.
However, I wanted to share with you my current results on WIDER.
In this table 11K represents 11,000 iteration model yolo-11K represents model with yolo.cfg anchors wider5-11K represents model with wider anchors that is computed for 416 (=13 * 32) input size (which could be found in this repo) wider5_hr-11K represents a model with the same anchors as in wider5-11K, but they are computed for 544(=17*32) input dims (so they are larger) voc-11K represents model with voc anchors.
As you can see from the table you can have 2-3% improvement with wider5_hr-11K However, this experiemnt did not finish yet, I want to try 19*32 anchors, I also want to try less number of anchors, for example 2 or 3 instead of 5. But I have problem with my validation code right now. Could you please share yours, how are u printing out wider detections in order to evaluate it using the code that WIDER provides? My added function is here: https://github.com/Jumabek/darknet/blob/master/src/detector.c#L249
@Jumabek Just saw this message. Sorry to reply you late.
I have not tested on wider dataset yet. My current evaluation is based on FDDB. Regarding your evaluation, it appears that the reason for the improvement is due to the increased resolution, rather than the switching of new anchors. I think it makes more sense to compare yolo-11k with wider5-11k?