Add tta to HTC and Cascade RCNN

[d0ng1ee]

Add tta to HTC and Cascade RCNN

write with the similar tta scheme of fasterrcnn in this repo

Here are the experiment results

Cascade Mask R-CNN

Backbone	Test method	Lr schd	Mem (GB)	Train time (s/iter)	Inf time (fps)	box AP	mask AP	Download
R-50-FPN	Orign repo	1x	5.3	0.683	7.4	41.2	35.7	model
R-50-FPN	TTA(hflip)					41.9	36.2
R-50-FPN	TTA(hflip+2scale)					42.7	36.8

Hybrid Task Cascade (HTC)

Backbone	Test method	Lr schd	Mem (GB)	Train time (s/iter)	Inf time (fps)	box AP	mask AP	Download
R-50-FPN	Orign repo	1x	7.4	0.936	4.1	42.1	37.3	model
R-50-FPN	TTA(hflip)					43.0	37.8
R-50-FPN	TTA(hflip+2scale)					43.6	38.4

experiment with the config file:

img_scale=[(1333, 800),(1600,960)],
flip=True,

[hellock]

Nice work! Could you fix the linting errors?

[d0ng1ee]

Nice work! Could you fix the linting errors?

I will try it. This is the first time I've contributed code to the open source community, and I've ignored the importance of the code format.

[clw5180]

Nice work! Could you fix the linting errors?

I will try it. This is the first time I've contributed code to the open source community, and I've ignored the importance of the code format.

Add tta to HTC and Cascade RCNN

write with the similar tta scheme of fasterrcnn in this repo

Here are the experiment results

Cascade Mask R-CNN

Backbone Test method Lr schd Mem (GB) Train time (s/iter) Inf time (fps) box AP mask AP Download
R-50-FPN Orign repo 1x 5.3 0.683 7.4 41.2 35.7 model
R-50-FPN TTA(hflip) 41.9 36.2
R-50-FPN TTA(hflip+2scale) 42.7 36.8

Hybrid Task Cascade (HTC)

Backbone Test method Lr schd Mem (GB) Train time (s/iter) Inf time (fps) box AP mask AP Download
R-50-FPN Orign repo 1x 7.4 0.936 4.1 42.1 37.3 model
R-50-FPN TTA(hflip) 43.0 37.8
R-50-FPN TTA(hflip+2scale) 43.6 38.4

experiment with the config file:

img_scale=[(1333, 800),(1600,960)],
flip=True,

Have you tried multi-scale training and how much mAP improvement you get ? Thanks！ @donglee-afar

[d0ng1ee]

Nice work! Could you fix the linting errors?

I will try it. This is the first time I've contributed code to the open source community, and I've ignored the importance of the code format.

Add tta to HTC and Cascade RCNN

write with the similar tta scheme of fasterrcnn in this repo

Here are the experiment results

Cascade Mask R-CNN

Backbone Test method Lr schd Mem (GB) Train time (s/iter) Inf time (fps) box AP mask AP Download
R-50-FPN Orign repo 1x 5.3 0.683 7.4 41.2 35.7 model
R-50-FPN TTA(hflip) 41.9 36.2
R-50-FPN TTA(hflip+2scale) 42.7 36.8

Hybrid Task Cascade (HTC)

Backbone Test method Lr schd Mem (GB) Train time (s/iter) Inf time (fps) box AP mask AP Download
R-50-FPN Orign repo 1x 7.4 0.936 4.1 42.1 37.3 model
R-50-FPN TTA(hflip) 43.0 37.8
R-50-FPN TTA(hflip+2scale) 43.6 38.4
experiment with the config file:

img_scale=[(1333, 800),(1600,960)],
flip=True,

Have you tried multi-scale training and how much mAP improvement you get ? Thanks！ @donglee-afar

I haven't try it on coco dataset, but in my experience on Kaggle and some other competition dataset, it usually can boost about 0.5~1.5% mAP

[mosheliv]

This looks promising... Will this work on retinanet? Is it only active as is in the "test" mode or also used with the API for single images?

[clw5180]

I haven't try it on coco dataset, but in my experience on Kaggle and some other competition dataset, it usually can boost about 0.5~1.5% mAP

Thanks a lot, I tried the multi-scale training but I find the mAP get worse. So do you think multi-scale training is necessary when I already use FPN ? @donglee-afar 另外能否留个企鹅啥的

[d0ng1ee]

qq:1955662900

[yhcao6]

Thanks for your work. This two statements seems redundant.

[yhcao6]

mask_feats seems redundant.

[d0ng1ee]

You are right, mask_feats does not need to be calculated here in cascadercnn because there is no semantic_feat to be add.
I have delete it and test it again, results is the same as what i posted before.