[Algorithm] Support H2RBox

[yangxue0827]

Motivation

• Support weakly supervised oriented object detection algorithm H2Rbox in MMRotate 1.x
• Support DOTA-v1.5 and DOTA-v2.0

Modification

• Create h2rbox.py at detectors
• Create h2rbox_head.py at dense_heads
• Create DOTAv15Dataset and DOTAv2Dataset in dota.py
• Create dotav15.py and dotav2.py at configs/base/datasets
• Create h2rbox_consistency_loss.py at losses
• Create README.md and metafile.yml at configs/h2rbox
• Create h2rbox-le90_r50_fpn_adamw-1x_dota.py , h2rbox-le90_r50_fpn_adamw-3x_dota.py and h2rbox-le90_r50_fpn_adamw-1x_dota-ms.py at configs/h2rbox
• Create h2rbox-le90_r50_fpn_adamw-1x_dotav15.py and h2rbox-le90_r50_fpn_adamw-3x_dotav15.py at configs/h2rbox/dotav15
• Create h2rbox-le90_r50_fpn_adamw-1x_dotav2.py and h2rbox-le90_r50_fpn_adamw-3x_dotav2.py at configs/h2rbox/dotav2
• Create h2rbox-le90_r50_fpn_adamw-1x_dior.py at configs/h2rbox/dior
• Create test_h2rbox at test_detectors
• Update test_loss.py at test_losses
• Update test_dota.py at test_datasets
• Update test_h2rbox_heads.py at test_dense_heads

Performance

• Performance on DOTA1.0

This PR

Backbone	mAP	Angle	lr schd	Mem (GB)	Inf Time (fps)	MS	Batch Size	Configs
ResNet50 (1024,1024,200)	68.75	le90	1x	6.25		-	2	h2rbox-le90_r50_fpn_adamw-1x_dota
ResNet50 (1024,1024,200)	71.31	le90	3x	6.64		-	2	h2rbox-le90_r50_fpn_adamw-3x_dota
ResNet50 (1024,1024,200)	74.43	le90	1x	6.12	-	-	2	h2rbox-le90_r50_fpn_adamw-1x_dota-ms

Original PyTorch Implementation (Paper)

Backbone	mAP	Angle	lr schd	Mem (GB)	Inf Time (fps)	MS	Batch Size	Configs
ResNet50 (1024,1024,200)	67.45	le90	1x	7.02	28.5	-	2	h2rbox_r50_adamw_fpn_1x_dota_le90
ResNet50 (1024,1024,200)	70.77	le90	3x	7.02	28.5	-	2	h2rbox_r50_adamw_fpn_3x_dota_le90
ResNet50 (1024,1024,200)	74.40	le90	1x	8.58	-	-	2	h2rbox_r50_adamw_fpn_3x_dota_ms_le90

Jittor Implementation

Backbone	mAP	Angle	lr schd	MS	Batch Size	Configs
ResNet50 (1024,1024,200)	67.62	le135	1x	-	2	h2rbox_obb_r50_adamw_fpn_1x_dota

• Performance on DOTA1.5

This PR

Backbone	mAP	Angle	lr schd	Mem (GB)	Inf Time (fps)	MS	Batch Size	Configs
ResNet50 (1024,1024,200)	60.33	le90	1x	6.83		-	2	h2rbox-le90_r50_fpn_adamw-1x_dotav15
ResNet50 (1024,1024,200)	62.65	le90	1x	6.34		-	2	h2rbox-le90_r50_fpn_adamw-3x_dotav15

Original PyTorch Implementation (Paper)

Backbone	mAP	Angle	lr schd	Mem (GB)	Inf Time (fps)	MS	Batch Size	Configs
ResNet50 (1024,1024,200)	60.19	le90	1x	10.68	25.8	-	2	h2rbox_r50_adamw_fpn_1x_dotav15_le90
ResNet50 (1024,1024,200)	62.60	le90	3x	10.68	25.8	-	2	h2rbox_r50_adamw_fpn_3x_dotav15_le90

• Performance on DOTA2.0

This PR

Backbone	mAP	Angle	lr schd	Mem (GB)	Inf Time (fps)	MS	Batch Size	Configs
ResNet50 (1024,1024,200)	47.08	le90	1x	7.58		-	2	h2rbox-le90_r50_fpn_adamw-1x_dotav2
ResNet50 (1024,1024,200)	50.20	le90	1x	7.74		-	2	h2rbox-le90_r50_fpn_adamw-3x_dotav2

Original PyTorch Implementation (Paper)

Backbone	mAP	Angle	lr schd	Mem (GB)	Inf Time (fps)	MS	Batch Size	Configs
ResNet50 (1024,1024,200)	45.87	le90	1x	11.57	25.0	-	2	h2rbox_r50_adamw_fpn_1x_dotav2_le90
ResNet50 (1024,1024,200)	47.86	le90	3x	11.57	25.0	-	2	h2rbox_r50_adamw_fpn_3x_dotav2_le90

• Performance on DIOR

This PR

Backbone	AP50:95	AP50	AP75	Angle	lr schd	Mem (GB)	Inf Time (fps)	MS	Batch Size	Configs
ResNet50 (1024,1024,200)	33.06	57.40	32.50	le90	1x	3.83		-	2	h2rbox-le90_r50_fpn_adamw-1x_dior

Original PyTorch Implementation (Paper)

Backbone	AP50:95	AP50	AP75	Angle	lr schd	Mem (GB)	Inf Time (fps)	MS	Batch Size	Configs
ResNet50 (1024,1024,200)	33.15	57.00	32.60	le90	1x	3.83		-	2	h2rbox_r50_adamw_fpn_1x_dior_le90

Notes

• MS means multiple scale image split.
• Inf Time was tested on a single RTX3090.
• The memory log of MMRotate 1.x is dynamic, so the last iteration is recorded here.

[codecov]

Codecov Report

Base: 61.20% // Head: 80.35% // Increases project coverage by +19.15% 🎉

Coverage data is based on head (b06e250) compared to base (e8c6eb8).
Patch coverage: 90.25% of modified lines in pull request are covered.

❗ Current head b06e250 differs from pull request most recent head 7a73f36. Consider uploading reports for the commit 7a73f36 to get more accurate results

Additional details and impacted files

@@             Coverage Diff              @@
##           dev-1.x     #644       +/-   ##
============================================
+ Coverage    61.20%   80.35%   +19.15%     
============================================
  Files           96       98        +2     
  Lines         6562     6806      +244     
  Branches       932      960       +28     
============================================
+ Hits          4016     5469     +1453     
+ Misses        2329     1024     -1305     
- Partials       217      313       +96     

Flag	Coverage Δ
unittests	?

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

Impacted Files	Coverage Δ
mmrotate/models/detectors/h2rbox.py	85.48% <85.48%> (ø)
mmrotate/models/dense_heads/h2rbox_head.py	89.83% <89.83%> (ø)
mmrotate/datasets/__init__.py	100.00% <100.00%> (ø)
mmrotate/datasets/dota.py	89.74% <100.00%> (+0.85%)	⬆️
mmrotate/models/dense_heads/__init__.py	100.00% <100.00%> (ø)
mmrotate/models/detectors/__init__.py	100.00% <100.00%> (ø)
mmrotate/models/losses/__init__.py	100.00% <100.00%> (ø)
mmrotate/models/losses/h2rbox_consistency_loss.py	100.00% <100.00%> (ø)
mmrotate/evaluation/metrics/dota_metric.py	87.00% <0.00%> (-1.13%)	⬇️
... and 28 more

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[zytx121]

LGTM

[liuyanyi]

LGTM