Experiment results of implement Cascade RCNN under Detectron.
Using ResNet50 as feature extractor, as well as 1x iterations.
Learning rate start as 0.01 (4GPU, 2 images per GPU).
Communication
All the experiments I have tried are shown as below, but the results are not as expected, any ideas and suggestions helpful
are welcomed.
Using statement
Folder detectron_cascade are codes to implement Cascade RCNN under Detectron, parallelizing with folder $Detectron/detcectron.
Folder configs/cascade/ contains yaml files conducting the Cascade RCNN model training.
MSCOCO experiments
mask iterative bbox rcnn results (using same IOU threshold in three stage of RCNN)
model is trained on coco2017train + val
experiments
dataset
box_ap
box_ap50
box_ap75
box_ap_small
box_ap_med
box_ap_large
mask_ap
mask_ap50
mask_ap75
mask_ap_small
mask_ap_med
mask_ap_large
mask-R50
test-dev
38.2%
60.05
41.5%
21.8%
40.3%
48.4%
34.3%
56.5%
36.3%
14.9%
36.1%
49.7%
cascade stage1
test-dev
38.3%
34.2%
cascade stage2
test-dev
38.9%
34.1%
cascade stage3
test-dev
38.9%
59.5%
42.1%
21.5%
40.7%
50.2%
34.0%
56.1%
35.9%
14.8%
35.5%
49.5%
cascade stage 1~2
test-dev
cascade stage 1~3
test-dev
mask cascade rcnn results beta version 1
(clip bbox and add invalid bbox check in DecodeBBoxOp)
model is trained on coco2017train + val
experiments
dataset
box_ap
box_ap50
box_ap75
box_ap_small
box_ap_med
box_ap_large
mask_ap
mask_ap50
mask_ap75
mask_ap_small
mask_ap_med
mask_ap_large
mask-R50
test-dev
38.2%
60.05
41.5%
21.8%
40.3%
48.4%
34.3%
56.5%
36.3%
14.9%
36.1%
49.7%
cascade stage1
test-dev
38.2%
59.9%
41.7%
21.7%
40.4%
48.4%
34.2%
56.4%
36.1%
15.0%
36.0%
49.5%
cascade stage2
test-dev
38.1%
58.5%
41.3%
18.2%
39.7%
53.4%
34.7%
56.5%
36.8%
15.1%
36.5%
50.6%
cascade stage3
test-dev
39.4%
57.5%
43.5%
21.4%
41.2%
51.1%
34.2%
55.0%
36.4%
14.7%
35.9%
49.9%
cascade stage 1~2
test-dev
cascade stage 1~3
test-dev
mask cascade rcnn results beta version 2
(screen out high iou boxes in DecodeBBoxOp)
model is trained on coco2017train
experiments
dataset
box_ap
box_ap50
box_ap75
box_ap_small
box_ap_med
box_ap_large
mask_ap
mask_ap50
mask_ap75
mask_ap_small
mask_ap_med
mask_ap_large
mask-R50
test-dev
38.6%
34.5%
cascade stage1
test-dev
cascade stage2
test-dev
cascade stage3
test-dev
39.06%
56.98%
43.28%
21.86%
41.54%
52.41%
34.20%
54.47%
36.65%
15.11%
36.47%
51.51%
cascade stage 1~2
test-dev
cascade stage 1~3
test-dev
mask cascade rcnn results beta version 3
(add weight to rcnn loss)
model is trained on coco2017train
experiments
dataset
box_ap
box_ap50
box_ap75
box_ap_small
box_ap_med
box_ap_large
mask_ap
mask_ap50
mask_ap75
mask_ap_small
mask_ap_med
mask_ap_large
mask-R50
test-dev
38.0%
34.5%
cascade stage1
test-dev
cascade stage2
test-dev
cascade stage3
test-dev
38.5%
57.2%
42.7%
20.9%
40.7%
49.1%
cascade stage 1~2
test-dev
cascade stage 1~3
test-dev
mask cascade rcnn results beta version 4
(use cls_agnostic_bbox_reg、specific lr_mult)
model is trained on coco2017train
experiments
dataset
box_ap
box_ap50
box_ap75
box_ap_small
box_ap_med
box_ap_large
mask_ap
mask_ap50
mask_ap75
mask_ap_small
mask_ap_med
mask_ap_large
mask-R50
test-dev(val)
38.00%(37.7%)
59.7%
41.3%
21.2%
40.2%
48.1%
34.20%(33.9%)
56.4%
36.0%
14.8%
36.0%
49.7%
cascade stage1
test-dev
36.8%
58.1%
40.0%
20.3%
39.0%
47.2%
33.5%
54.9%
35.4%
14.3%
35.2%
48.2%
cascade stage2
test-dev
38.9%
58.6%
42.8%
21.0%
40.9%
50.5%
34.4%
55.6%
36.6%
14.5%
36.0%
50.2%
cascade stage3
test-dev
38.9%
57.4%
43.1%
20.8%
40.8%
51.0%
34.3%
54.7%
36.7%
14.4%
35.8%
50.0%
cascade stage 1~2
test-dev
38.9%
59.0%
42.7%
21.3%
41.0%
50.5%
34.4%
55.8%
36.5%
14.6%
36.0%
50.3%
cascade stage 1~3
test-dev(val)
39.50%(39.14%)
58.90%(58.36%)
43.40%(42.85%)
21.50%(21.41%)
41.40%(41.52%)
51.30%(53.03%)
34.60%(34.37%)
55.80%(55.22%)
36.80%(36.57%)
14.80%(15.17%)
36.20%(36.5%)
50.40%(52.09%)
mask cascade rcnn results beta version 4 large iter
model is trained on coco2017train, learning rate start at 0.01, reduce to 0.001 at 160000 iterations and 0.0001 at 240000 iterations
experiments
dataset
box_ap
box_ap50
box_ap75
box_ap_small
box_ap_med
box_ap_large
mask_ap
mask_ap50
mask_ap75
mask_ap_small
mask_ap_med
mask_ap_large
cascade stage 1~3
test-dev(val)
40.10%(39.75%)
59.40%(58.91%)
43.90%(43.56%)
22.00%(21.78%)
41.90%(42.13%)
51.90%(54.24%)
35.00%(34.73%)
56.30%(55.82%)
37.20%(36.90%)
15.10%(14.85%)
36.60%(36.93%)
51.00%(53.20%)
faster cascade rcnn results
model is trained on coco2017train
experiments
dataset
box_ap
box_ap50
box_ap75
box_ap_small
box_ap_med
box_ap_large
faster-FPN-R50
test-dev(val)
(36.7%)
(58.45%)
(39.61%)
(21.12%)
(39.85%)
(48.13%)
cascade stage1
test-dev(val)
cascade stage2
test-dev(val)
cascade stage3
test-dev(val)
cascade stage 1~2
test-dev(val)
cascade stage 1~3
test-dev(val)
(37.31%)
(55.51%)
(40.65%)
(20.30%)
(39.87%)
(49.21%)
PASCAL VOC experiments
model is trained on voc0712 trainval, tested on voc2007 test, using coco evaluation metrics