[WIP] CFA & Oriented Reppoints & SASM

README.md

@@ -118,6 +118,9 @@ python run_net.py --config-file=configs/base.py --task=test
 | RSDet-R50-FPN | DOTA1.0|1024/200|Flip|-| SGD | 1x | 68.41 | [arxiv](https://arxiv.org/abs/1911.08299) | [config](configs/rotated_retinanet/rsdet_obb_r50_fpn_1x_dota_lmr5p.py) | [model](https://cloud.tsinghua.edu.cn/f/642e200f5a8a420eb726/?dl=1) |
 | ATSS-R50-FPN|DOTA1.0|1024/200| flip|-| SGD | 1x | 72.44 | [arxiv](https://arxiv.org/abs/1912.02424) | [config](configs/rotated_retinanet/rotated_retinanet_obb_r50_fpn_1x_dota_atss.py) | [model](https://cloud.tsinghua.edu.cn/f/5168189dcd364eaebce5/?dl=1) |
 | Reppoints-R50-FPN|DOTA1.0|1024/200| flip|-| SGD | 1x | 56.34 | [arxiv](https://arxiv.org/abs/1904.11490) | [config](configs/rotated_retinanet/rotated_retinanet_obb_r50_fpn_1x_dota_atss.py) | [model](https://cloud.tsinghua.edu.cn/f/be359ac932c84f9c839e/?dl=1) |
+| CFA-R50-FPN|DOTA1.0|1024/200| flip|-| SGD | 1x |  | [arxiv]() | [config]() | [model]() |
+| Oriented-Reppoints-R50-FPN|DOTA1.0|1024/200| flip|-| SGD | 1x |  | [arxiv]() | [config]() | [model]() |
+| SASM-R50-FPN|DOTA1.0|1024/200| flip|-| SGD | 1x |  | [arxiv]() | [config]() | [model]() |
 
 
 **Notice**:
@@ -153,9 +156,11 @@ python run_net.py --config-file=configs/base.py --task=test
 - :heavy_check_mark: Reppoints
 - :heavy_check_mark: RSDet
 - :heavy_check_mark: ATSS
+- :heavy_check_mark: CFA
+- :heavy_check_mark: Oriented Reppoints
+- :heavy_check_mark: SASM
 - :clock3: R3Det
 - :clock3: Cascade R-CNN
-- :clock3: Oriented Reppoints
 - :heavy_plus_sign: DCL
 - :heavy_plus_sign: Double Head OBB
 - :heavy_plus_sign: Guided Anchoring

configs/cfa_r50_fpn_1x_dota.py

@@ -0,0 +1,161 @@
+# model settings
+model = dict(
+    type='RotatedRetinaNet',
+    backbone=dict(
+        type='Resnet50',
+        frozen_stages=1,
+        return_stages=["layer1","layer2","layer3","layer4"],
+        pretrained= True),
+    neck=dict(
+        type='FPN',
+        in_channels=[256, 512, 1024, 2048],
+        out_channels=256,
+        start_level=1,
+        add_extra_convs="on_input",
+        norm_cfg = dict(type='GN', num_groups=32),
+        num_outs=5),
+    bbox_head=dict(
+        type='RotatedRepPointsHead',
+        num_classes=15,
+        in_channels=256,
+        feat_channels=256,
+        point_feat_channels=256,
+        stacked_convs=3,
+        num_points=9,
+        gradient_mul=0.3,
+        point_strides=[8, 16, 32, 64, 128],
+        point_base_scale=2,
+        norm_cfg = dict(type='GN', num_groups=32),
+        loss_cls=dict(
+            type='FocalLoss',
+            use_sigmoid=True,
+            gamma=2.0,
+            alpha=0.25,
+            loss_weight=1.0),
+        loss_bbox_init=dict(type='ConvexGIoULoss', loss_weight=0.375),
+        loss_bbox_refine=dict(type='ConvexGIoULoss', loss_weight=1.0),
+        transform_method='rotrect',
+        use_reassign=True,
+        topk=6,
+        anti_factor=0.75,
+        train_cfg=dict(
+            init=dict(
+                assigner=dict(type='ConvexAssigner', scale=4, pos_num=1, assigned_labels_filled=-1),
+                allowed_border=-1,
+                pos_weight=-1,
+                debug=False),
+            refine=dict(
+                assigner=dict(
+                    type='MaxConvexIoUAssigner',
+                    pos_iou_thr=0.1,
+                    neg_iou_thr=0.1,
+                    min_pos_iou=0,
+                    ignore_iof_thr=-1,
+                    assigned_labels_filled=-1,
+                    iou_calculator=dict(type='ConvexOverlaps')),
+                allowed_border=-1,
+                pos_weight=-1,
+                debug=False)),
+        test_cfg=dict(
+            nms_pre=2000,
+            min_bbox_size=0,
+            score_thr=0.05,
+            nms=dict(iou_thr=0.4),
+            max_per_img=2000))
+    )
+dataset = dict(
+    train=dict(
+        type="DOTADataset",
+        dataset_dir='/home/zytx121/mmrotate/data/processed_DOTA/trainval_1024_200_1.0',
+        transforms=[
+            dict(
+                type="RotatedResize",
+                min_size=1024,
+                max_size=1024
+            ),
+            dict(type='RotatedRandomFlip', prob=0.5, direction="horizontal"),
+            dict(type='RotatedRandomFlip', prob=0.5, direction="vertical"),
+            dict(
+                type = "Pad",
+                size_divisor=32),
+            dict(
+                type = "Normalize",
+                mean =  [123.675, 116.28, 103.53],
+                std = [58.395, 57.12, 57.375],
+                to_bgr=False,)
+
+        ],
+        batch_size=1,
+        num_workers=4,
+        shuffle=True,
+        filter_empty_gt=False
+    ),
+    val=dict(
+        type="DOTADataset",
+        dataset_dir='/home/zytx121/mmrotate/data/processed_DOTA/trainval_1024_200_1.0',
+        transforms=[
+            dict(
+                type="RotatedResize",
+                min_size=1024,
+                max_size=1024
+            ),
+            dict(
+                type = "Pad",
+                size_divisor=32),
+            dict(
+                type = "Normalize",
+                mean =  [123.675, 116.28, 103.53],
+                std = [58.395, 57.12, 57.375],
+                to_bgr=False),
+        ],
+        batch_size=2,
+        num_workers=4,
+        shuffle=False
+    ),
+    test=dict(
+        type="ImageDataset",
+        images_dir='/home/zytx121/mmrotate/data/processed_DOTA/test_1024_200_1.0/images',
+        transforms=[
+            dict(
+                type="RotatedResize",
+                min_size=1024,
+                max_size=1024
+            ),
+            dict(
+                type = "Pad",
+                size_divisor=32),
+            dict(
+                type = "Normalize",
+                mean =  [123.675, 116.28, 103.53],
+                std = [58.395, 57.12, 57.375],
+                to_bgr=False,),
+        ],
+        num_workers=4,
+        batch_size=1,
+    )
+)
+
+optimizer = dict(
+    type='SGD',  
+    lr=0.008,
+    momentum=0.9, 
+    weight_decay=0.0001,
+    grad_clip=dict(
+        max_norm=35, 
+        norm_type=2))
+
+scheduler = dict(
+    type='StepLR',
+    warmup='linear',
+    warmup_iters=500,
+    warmup_ratio=1.0 / 3,
+    milestones=[7, 10])
+
+
+logger = dict(
+    type="RunLogger")
+
+max_epoch = 12
+eval_interval = 1
+checkpoint_interval = 1
+log_interval = 50

configs/sasm_obb_r50_fpn_1x_dota.py

@@ -0,0 +1,155 @@
+# model settings
+model = dict(
+    type='RotatedRetinaNet',
+    backbone=dict(
+        type='Resnet50',
+        frozen_stages=1,
+        return_stages=["layer1","layer2","layer3","layer4"],
+        pretrained= True),
+    neck=dict(
+        type='FPN',
+        in_channels=[256, 512, 1024, 2048],
+        out_channels=256,
+        start_level=1,
+        add_extra_convs="on_input",
+        norm_cfg = dict(type='GN', num_groups=32),
+        num_outs=5),
+    bbox_head=dict(
+        type='SAMRepPointsHead',
+        num_classes=15,
+        in_channels=256,
+        feat_channels=256,
+        point_feat_channels=256,
+        stacked_convs=3,
+        num_points=9,
+        gradient_mul=0.3,
+        point_strides=[8, 16, 32, 64, 128],
+        point_base_scale=2,
+        norm_cfg = dict(type='GN', num_groups=32),
+        loss_cls=dict(
+            type='FocalLoss',
+            use_sigmoid=True,
+            gamma=2.0,
+            alpha=0.25,
+            loss_weight=1.0),
+        loss_bbox_init=dict(type='ConvexGIoULoss', loss_weight=0.375),
+        loss_bbox_refine=dict(type='BCConvexGIoULoss', loss_weight=1.0),
+        transform_method='rotrect',
+        use_reassign=False,
+        topk=6,
+        anti_factor=0.75,
+        train_cfg=dict(
+            init=dict(
+                assigner=dict(type='ConvexAssigner', scale=4, pos_num=1, assigned_labels_filled=-1),
+                allowed_border=-1,
+                pos_weight=-1,
+                debug=False),
+            refine=dict(
+                assigner=dict(type='SASAssigner', topk=9),
+                allowed_border=-1,
+                pos_weight=-1,
+                debug=False)),
+        test_cfg=dict(
+            nms_pre=2000,
+            min_bbox_size=0,
+            score_thr=0.05,
+            nms=dict(iou_thr=0.4),
+            max_per_img=2000))
+    )
+dataset = dict(
+    train=dict(
+        type="DOTADataset",
+        dataset_dir='/home/zytx121/mmrotate/data/processed_DOTA/trainval_1024_200_1.0',
+        transforms=[
+            dict(
+                type="RotatedResize",
+                min_size=1024,
+                max_size=1024
+            ),
+            dict(type='RotatedRandomFlip', prob=0.5, direction="horizontal"),
+            dict(type='RotatedRandomFlip', prob=0.5, direction="vertical"),
+            dict(
+                type = "Pad",
+                size_divisor=32),
+            dict(
+                type = "Normalize",
+                mean =  [123.675, 116.28, 103.53],
+                std = [58.395, 57.12, 57.375],
+                to_bgr=False,)
+
+        ],
+        batch_size=2,
+        num_workers=4,
+        shuffle=True,
+        filter_empty_gt=False
+    ),
+    val=dict(
+        type="DOTADataset",
+        dataset_dir='/home/zytx121/mmrotate/data/processed_DOTA/trainval_1024_200_1.0',
+        transforms=[
+            dict(
+                type="RotatedResize",
+                min_size=1024,
+                max_size=1024
+            ),
+            dict(
+                type = "Pad",
+                size_divisor=32),
+            dict(
+                type = "Normalize",
+                mean =  [123.675, 116.28, 103.53],
+                std = [58.395, 57.12, 57.375],
+                to_bgr=False),
+        ],
+        batch_size=2,
+        num_workers=4,
+        shuffle=False
+    ),
+    test=dict(
+        type="ImageDataset",
+        images_dir='/home/zytx121/mmrotate/data/processed_DOTA/test_1024_200_1.0/images',
+        transforms=[
+            dict(
+                type="RotatedResize",
+                min_size=1024,
+                max_size=1024
+            ),
+            dict(
+                type = "Pad",
+                size_divisor=32),
+            dict(
+                type = "Normalize",
+                mean =  [123.675, 116.28, 103.53],
+                std = [58.395, 57.12, 57.375],
+                to_bgr=False,),
+        ],
+        num_workers=4,
+        batch_size=1,
+    )
+)
+
+optimizer = dict(
+    type='SGD', 
+    # lr=0.01/4., #0.0,#0.01*(1/8.), 
+    lr=0.008,
+    momentum=0.9, 
+    weight_decay=0.0001,
+    grad_clip=dict(
+        max_norm=35, 
+        norm_type=2))
+
+scheduler = dict(
+    type='StepLR',
+    warmup='linear',
+    warmup_iters=500,
+    warmup_ratio=1.0 / 3,
+    milestones=[7, 10])
+
+
+logger = dict(
+    type="RunLogger")
+
+max_epoch = 12
+eval_interval = 1
+checkpoint_interval = 1
+log_interval = 50

projects/oriented_reppoints/README.md

@@ -0,0 +1,45 @@
+## Oriented RepPoints
+> [Oriented RepPoints for Aerial Object Detection](https://openaccess.thecvf.com/content/CVPR2022/papers/Li_Oriented_RepPoints_for_Aerial_Object_Detection_CVPR_2022_paper.pdf)
+
+<!-- [ALGORITHM] -->
+### Abstract
+
+<div align=center>
+<img src="https://raw.githubusercontent.com/zytx121/image-host/main/imgs/oriented_reppoints.png" width="800"/>
+</div>
+
+In contrast to the generic object, aerial targets are often non-axis aligned with arbitrary orientations having
+the cluttered surroundings. Unlike the mainstreamed approaches regressing the bounding box orientations, this paper
+proposes an effective adaptive points learning approach to aerial object detection by taking advantage of the adaptive
+points representation, which is able to capture the geometric information of the arbitrary-oriented instances.
+To this end, three oriented conversion functions are presented to facilitate the classification and localization
+with accurate orientation. Moreover, we propose an effective quality assessment and sample assignment scheme for
+adaptive points learning toward choosing the representative oriented reppoints samples during training, which is
+able to capture the non-axis aligned features from adjacent objects or background noises. A spatial constraint is
+introduced to penalize the outlier points for roust adaptive learning. Experimental results on four challenging
+aerial datasets including DOTA, HRSC2016, UCAS-AOD and DIOR-R, demonstrate the efficacy of our proposed approach.
+
+### Training
+```sh
+python run_net.py --config-file=configs/oriented_reppoints_r50_fpn_1x_dota.py --task=train
+```
+
+### Testing
+```sh
+python run_net.py --config-file=configs/oriented_reppoints_r50_fpn_1x_dota.py --task=test
+```
+
+### Performance
+|   Models    | Dataset| Sub_Image_Size/Overlap |Train Aug | Test Aug | Optim | Lr schd | mAP    | Paper |                             Config                             | Download   |
+|:-----------:| :-----: |:-----:|:-----:| :-----: | :-----:| :-----:| :----: |:--------:|:--------------------------------------------------------------:| :--------: |
+| Oriented-Reppoints-R50-FPN | DOTA1.0|1024/200| flip|-|  SGD   |   1x    | 66.99   | [paper](https://openaccess.thecvf.com/content/CVPR2022/papers/Li_Oriented_RepPoints_for_Aerial_Object_Detection_CVPR_2022_paper.pdf)| [config](configs/oriented_reppoints_r50_fpn_1x_dota.py) | [model]() |
+
+### Citation
+```
+@inproceedings{li2022ori,
+    title={Oriented RepPoints for Aerial Object Detection},
+    author={Wentong Li, Yijie Chen, Kaixuan Hu, Jianke Zhu},
+    booktitle={Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
+    year={2022}
+}
+```