Higher HRNet multi-scale results on CrowdPose question

[nibox]

Hi,
first of all, thank you very much for reproducing the results on CrowdPose and publishing your model.
My question is do you have results for HigherHRNet w32 on the CrowdPose test split using multi-scale inference?
I tried to produce them by setting test_scale_factor=[2.0, 1.0, 0.5] in the val_pipline config, but the results are worse compared to single-scale inference, so I am not sure if I have done everything correct.

[jin-s13]

Yes, you just need to set test_scale_factor in the val_pipline config.
I have tried the test_scale_factor=[2, 1], and the results look good.

Average Precision (AP) @[ IoU=0.50:0.95 | type= all | maxDets= 20 ] = 0.661
Average Precision (AP) @[ IoU=0.50 | type= all | maxDets= 20 ] = 0.864
Average Precision (AP) @[ IoU=0.75 | type= all | maxDets= 20 ] = 0.710
Average Recall (AR) @[ IoU=0.50:0.95 | type= all | maxDets= 20 ] = 0.734
Average Recall (AR) @[ IoU=0.50 | type= all | maxDets= 20 ] = 0.912
Average Recall (AR) @[ IoU=0.75 | type= all | maxDets= 20 ] = 0.778
Average Precision (AP) @[ IoU=0.50:0.95 | type= easy | maxDets= 20 ] = 0.742
Average Precision (AP) @[ IoU=0.50:0.95 | type=medium | maxDets= 20 ] = 0.670
Average Precision (AP) @[ IoU=0.50:0.95 | type= hard | maxDets= 20 ] = 0.566

I will try test_scale_factor=[2.0, 1.0, 0.5] today. Will let you know when I get the results.

[jin-s13]

I have just tried test_scale_factor=[2.0, 1.0, 0.5]. The results are listed below.

Average Precision (AP) @[ IoU=0.50:0.95 | type= all | maxDets= 20 ] = 0.614
Average Precision (AP) @[ IoU=0.50 | type= all | maxDets= 20 ] = 0.801
Average Precision (AP) @[ IoU=0.75 | type= all | maxDets= 20 ] = 0.656
Average Recall (AR) @[ IoU=0.50:0.95 | type= all | maxDets= 20 ] = 0.710
Average Recall (AR) @[ IoU=0.50 | type= all | maxDets= 20 ] = 0.895
Average Recall (AR) @[ IoU=0.75 | type= all | maxDets= 20 ] = 0.748
Average Precision (AP) @[ IoU=0.50:0.95 | type= easy | maxDets= 20 ] = 0.699
Average Precision (AP) @[ IoU=0.50:0.95 | type=medium | maxDets= 20 ] = 0.627
Average Precision (AP) @[ IoU=0.50:0.95 | type= hard | maxDets= 20 ] = 0.501

It seems that including 0.5 in the test scales is not always beneficial.

[nibox]

Thank you very much. For test_scale_factor=[2.0, 1.0, 0.5] i get the same results.
Now I am wondering which configuration of scales factors the original authors used.