Implementation of TridentNet

[Chrisfsj2051]

Implementation of TridentNet.

Please note that this implementation refers to the implementation in detectron2, which is slightly different from the origin paper. In paper, the author use r50 to conduct abaltion study, but report final results based on r101. While in detectron2, all experiments are conducted with r50. Besides, the author use all 3 branches to yield best performance, but detectron2 only uses the middle branch (index=1, dilation=2).

This implementation is still work in progress, for slightly lower accuracy comparing to detectron2. The unit test will be added after accuracy meets expect.

backbone	schedule	mAP-mmdet	mAP-detr2
r50	1x	37.8	38.0
r50	3x	40.3	40.6
r101	3x	Out of Memory	43.6

Update 7.20: The faster-rcnn C4 baseline with 3x schedule failed to reach the accuracy in detectron2.

type	schedule	mAP-mmdet	mAP-detr2
frcnn-c4-r50	3x	37.8	38.4

[hezhu1996]

Hi @Chrisfsj2051 , I'm very interested in TridentNet as well. Is it runable in your repository?

[Chrisfsj2051]

Hi @TWDH , Sorry for the Late Reply. Sure, it is already runnable.

[codecov]

Codecov Report

Merging #3313 (6a0862b) into master (fb287b1) will decrease coverage by 0.07%.
The diff coverage is 61.84%.

@@            Coverage Diff             @@
##           master    #3313      +/-   ##
==========================================
- Coverage   62.66%   62.59%   -0.08%     
==========================================
  Files         232      235       +3     
  Lines       17597    17818     +221     
  Branches     2942     2983      +41     
==========================================
+ Hits        11028    11153     +125     
- Misses       6031     6123      +92     
- Partials      538      542       +4     

Flag	Coverage Δ
unittests	62.56% <61.84%> (-0.07%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
...mdet/core/bbox/iou_calculators/iou2d_calculator.py	96.77% <ø> (ø)
mmdet/datasets/pipelines/formating.py	66.37% <ø> (ø)
mmdet/models/backbones/resnet.py	99.30% <ø> (ø)
mmdet/models/dense_heads/gfl_head.py	27.35% <ø> (ø)
mmdet/models/dense_heads/yolo_head.py	34.04% <ø> (ø)
mmdet/models/necks/pafpn.py	39.58% <0.00%> (-3.60%)	⬇️
mmdet/models/roi_heads/trident_roi_head.py	23.63% <23.63%> (ø)
mmdet/core/post_processing/bbox_nms.py	85.96% <33.33%> (-3.13%)	⬇️
mmdet/models/detectors/trident_faster_rcnn.py	43.33% <43.33%> (ø)
mmdet/datasets/samplers/distributed_sampler.py	33.33% <50.00%> (+3.92%)	⬆️
... and 11 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update fb287b1...95f6287. Read the comment docs.

[Chrisfsj2051]

Hi @Johnson-Wang and @hellock ,

I tried to train TridentNet (as well as FRCNN-C4) with both Iteration-Based and Epoch-Based to see the difference, here's the result:

setting	x-based	schedule	mAP-mmdet
mmdet-trident	epoch	1x	37.8
mmdet-trident	iter	1x	37.9
detr2-trident	iter	1x	38.0
mmdet-trident	epoch	3x	42.2
mmdet-trident	iter	3x	42.1
detr2-trident	iter	3x	42.6
mmdet-FrcnnC4	epoch	3x	38.1
mmdet-FrcnnC4	iter	3x	38.1
detr2-FrcnnC4	iter	3x	38.4

As shown in this table, with iteration-based 1x schedule (~12.3 epoch), this implementation yield 37.9mAP (0.1 lower than detr2). However, under 3x schedule, TridentNet is ~0.4 lower than detr2 and Frcnn-C4 is ~0.3 lower than detr2.

I'm wondering if it is qualified to be viewed or merged?

[hezhu1996]

Hi @TWDH , Sorry for the Late Reply. Sure, it is already runnable.

@Chrisfsj2051 Hi, Thanks for reply. I have a question: What's your baseline's mAP? As far as I know, the official FasterRCNN in mmedetection is already 37.8% at https://github.com/open-mmlab/mmdetection/tree/master/configs/faster_rcnn. detectron2 improves the performance from 35.7 to 38. Does you code have the same improvement? Thanks, just curious.

[Chrisfsj2051]

Hi @TWDH , Sorry for the Late Reply. Sure, it is already runnable.

@Chrisfsj2051 Hi, Thanks for reply. I have a question: What's your baseline's mAP? As far as I know, the official FasterRCNN in mmedetection is already 37.8% at https://github.com/open-mmlab/mmdetection/tree/master/configs/faster_rcnn. detectron2 improves the performance from 35.7 to 38. Does you code have the same improvement? Thanks, just curious.

Hi @TWDH , the 35.7 mAP FRCNN in detectron2 is FRCNN-C4. The 37.8 mAP FRCNN in mmdet that you mentioned above, however, is FRCNN-FPN, which has 37.9 mAP in detectron2 implementation (R50-FPN). TridentNet is based on FRCNN C4.

[hellock]

Task linked: CU-4btyv9 TridentNet PR review & Merge

[Chrisfsj2051]

Task linked: CU-4btyv9 TridentNet PR review & Merge

It shows "You do not have access to this task code: ACCESS_083" with account "fangshijie@sensetime.com"

[CuongNN218]

@Chrisfsj2051 I'm really interested in your work, especially ResNet101-deformable. How can i use your work? I'm a github novice. thanks

[Chrisfsj2051]

Using all of 3 branches for inference, we can improve the mAP of mstrain-3x setting model from 40.3 to 40.6.