evaluating widerFace validation datesets mAP problem

[kuaikuaikim]

Hi, @dechunwang I try to evaluate widerFace validation datasets(easy medium hard) with your pretrained mode, and just get this mAP(0.910, 0.897, 0.620) VS origin SSH(0.919, 0.907, 0.814)。Why the hard sets score is less than the origin SSH so much. My trained model get the same score like yours. Does the generate target anchor layer still have the problem?

My Fixed parameter and Method:

1. resize the shortest side to 1200
2. thres is 0.005(Detections with a probability less than this threshold are ignored)

[dechunwang]

Hi, The threshold for evaluation should be 0.05 instead of 0.005. Anchor target problem was fixed. Are you using eval.py from this repo?

[kuaikuaikim]

@dechunwang . sorry my clerical error, i used the 0.05, but get this mAP(0.910, 0.897, 0.620) VS origin SSH(0.919, 0.907, 0.814).

[dechunwang]

I run the eval.py with the pre-trained model and I can get the correct result.

wider_pr_curve_int_easy_val.pdf wider_pr_curve_int_hard_val.pdf wider_pr_curve_int_medium_val.pdf Mayebe you should delete some baseline at model/wider_eval_tools/plot/baselines/Val/setting_int/ folder, just in case if you mistakely looked at other baselines.

[dechunwang]

You can double-check the anchor layer and try to retrain the model. Since I am currently working on another project, I might not able to look into it.