Code Release of ECCV 2020 Spotlight paper for Side-Aware Boundary Loc…

…alization for More Precise Object Detection (open-mmlab#3603) * add sabl two stage * add sabl retina * ret cfg bug fix * test bug fix * minor update * update * add r101 two stage * update * add cfgs * add cfgs * update cfgs * format * format * add readme * fix isort * update * update readme * add doc string for sabl retina head * add doc string and rename some functions * add docstring for bucketing coder * update docstring * bucket_num -> num_buckets * bucket_pw -> bucket_w bucket_ph -> bucket_h * update label2onehot * update bucketing bbox coder doc * update * typo fix * bboxes_ -> rescaled_bboxes * rename some params in sabl head * init with mmcv.cnn * update doc * rename pos->post * update cfgs * update test cfg * update * add unitest for sabl head * add unitest for sabl retina * rename * minor rename * minor update * update docstring * update * use F.one_hot * update docstring * update test heads * update ReadMe * fix
gouthamkallempudi · Sep 1, 2020 · 26562a1 · 26562a1
1 parent 08d1402
commit 26562a1
Show file tree

Hide file tree

Showing 22 changed files with 2,365 additions and 19 deletions.
diff --git a/README.md b/README.md
@@ -97,6 +97,7 @@ Supported methods:
 - [x] [DetectoRS](configs/detectors/README.md)
 - [x] [Generalized Focal Loss](configs/gfl/README.md)
 - [x] [CornerNet](configs/cornernet/README.md)
+- [x] [Side-Aware Boundary Localization](configs/sabl/README.md)
 - [x] [YOLOv3](configs/yolo/README.md)
 
 Some other methods are also supported in [projects using MMDetection](./docs/projects.md).

diff --git a/configs/sabl/README.md b/configs/sabl/README.md
@@ -0,0 +1,36 @@
+# Side-Aware Boundary Localization for More Precise Object Detection
+
+## Introduction
+
+We provide config files to reproduce the object detection results in the ECCV 2020 Spotlight paper for [Side-Aware Boundary Localization for More Precise Object Detection](https://arxiv.org/abs/1912.04260).
+
+```
+@inproceedings{Wang_2020_ECCV,
+    title = {Side-Aware Boundary Localization for More Precise Object Detection},
+    author = {Wang, Jiaqi and Zhang, Wenwei and Cao, Yuhang and Chen, Kai and Pang, Jiangmiao and Gong, Tao and Shi, Jianping, Loy, Chen Change and Lin, Dahua},
+    booktitle = {ECCV},
+    year = {2020}
+}
+```
+
+## Results and Models
+
+The results on COCO 2017 val is shown in the below table. (results on test-dev are usually slightly higher than val).
+Single-scale testing (1333x800) is adopted in all results.
+
+
+|       Method       | Backbone  | Lr schd | ms-train | box AP |                                                                                                                                                        Download                                                                                                                                                         |
+| :----------------: | :-------: | :-----: | :------: | :----: | :---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: |
+| SABL Faster R-CNN  | R-50-FPN  |   1x    |    N     |  39.9  |    [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_faster_rcnn_r50_fpn_1x_coco/sabl_faster_rcnn_r50_fpn_1x_coco-e867595b.pth) &#124; [log](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_faster_rcnn_r50_fpn_1x_coco/20200830_130324.log.json)    |
+| SABL Faster R-CNN  | R-101-FPN |   1x    |    N     |  41.7  |  [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_faster_rcnn_r101_fpn_1x_coco/sabl_faster_rcnn_r101_fpn_1x_coco-f804c6c1.pth) &#124; [log](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_faster_rcnn_r101_fpn_1x_coco/20200830_183949.log.json)   |
+| SABL Cascade R-CNN | R-50-FPN  |   1x    |    N     |  41.6  |  [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_cascade_rcnn_r50_fpn_1x_coco/sabl_cascade_rcnn_r50_fpn_1x_coco-e1748e5e.pth) &#124; [log](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_cascade_rcnn_r50_fpn_1x_coco/20200831_033726.log.json)   |
+| SABL Cascade R-CNN | R-101-FPN |   1x    |    N     |  43.0  | [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_cascade_rcnn_r101_fpn_1x_coco/sabl_cascade_rcnn_r101_fpn_1x_coco-2b83e87c.pth) &#124; [log](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_cascade_rcnn_r101_fpn_1x_coco/20200831_141745.log.json) |
+
+|     Method     | Backbone  |  GN   | Lr schd |  ms-train   | box AP |                                                                                                                                                                         Download                                                                                                                                                                         |
+| :------------: | :-------: | :---: | :-----: | :---------: | :----: | :------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: |
+| SABL RetinaNet | R-50-FPN  |   N   |   1x    |      N      |  37.7  |                       [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_retinanet_r50_fpn_1x_coco/sabl_retinanet_r50_fpn_1x_coco-6c54fd4f.pth) &#124; [log](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_retinanet_r50_fpn_1x_coco/20200830_053451.log.json)                        |
+| SABL RetinaNet | R-50-FPN  |   Y   |   1x    |      N      |  38.8  |                   [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_retinanet_r50_fpn_gn_1x_coco/sabl_retinanet_r50_fpn_gn_1x_coco-e16dfcf1.pth) &#124; [log](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_retinanet_r50_fpn_gn_1x_coco/20200831_141955.log.json)                   |
+| SABL RetinaNet | R-101-FPN |   N   |   1x    |      N      |  39.7  |                      [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_retinanet_r101_fpn_1x_coco/sabl_retinanet_r101_fpn_1x_coco-42026904.pth) &#124; [log](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_retinanet_r101_fpn_1x_coco/20200831_034256.log.json)                      |
+| SABL RetinaNet | R-101-FPN |   Y   |   1x    |      N      |  40.5  |                 [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_retinanet_r101_fpn_gn_1x_coco/sabl_retinanet_r101_fpn_gn_1x_coco-40a893e8.pth) &#124; [log](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_retinanet_r101_fpn_gn_1x_coco/20200830_201422.log.json)                  |
+| SABL RetinaNet | R-101-FPN |   Y   |   2x    | Y (640~800) |  42.9  | [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_retinanet_r101_fpn_gn_2x_ms_640_800_coco/sabl_retinanet_r101_fpn_gn_2x_ms_640_800_coco-1e63382c.pth) &#124; [log](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_retinanet_r101_fpn_gn_2x_ms_640_800_coco/20200830_144807.log.json) |
+| SABL RetinaNet | R-101-FPN |   Y   |   2x    | Y (480~960) |  43.6  | [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_retinanet_r101_fpn_gn_2x_ms_480_960_coco/sabl_retinanet_r101_fpn_gn_2x_ms_480_960_coco-5342f857.pth) &#124; [log](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/v2.0/sabl/sabl_retinanet_r101_fpn_gn_2x_ms_480_960_coco/20200830_164537.log.json) |
diff --git a/configs/sabl/sabl_cascade_rcnn_r101_fpn_1x_coco.py b/configs/sabl/sabl_cascade_rcnn_r101_fpn_1x_coco.py
@@ -0,0 +1,88 @@
+_base_ = [
+    '../_base_/models/cascade_rcnn_r50_fpn.py',
+    '../_base_/datasets/coco_detection.py',
+    '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py'
+]
+# model settings
+model = dict(
+    pretrained='torchvision://resnet101',
+    backbone=dict(depth=101),
+    roi_head=dict(bbox_head=[
+        dict(
+            type='SABLHead',
+            num_classes=80,
+            cls_in_channels=256,
+            reg_in_channels=256,
+            roi_feat_size=7,
+            reg_feat_up_ratio=2,
+            reg_pre_kernel=3,
+            reg_post_kernel=3,
+            reg_pre_num=2,
+            reg_post_num=1,
+            cls_out_channels=1024,
+            reg_offset_out_channels=256,
+            reg_cls_out_channels=256,
+            num_cls_fcs=1,
+            num_reg_fcs=0,
+            reg_class_agnostic=True,
+            norm_cfg=None,
+            bbox_coder=dict(
+                type='BucketingBBoxCoder', num_buckets=14, scale_factor=1.7),
+            loss_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
+            loss_bbox_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+            loss_bbox_reg=dict(type='SmoothL1Loss', beta=0.1,
+                               loss_weight=1.0)),
+        dict(
+            type='SABLHead',
+            num_classes=80,
+            cls_in_channels=256,
+            reg_in_channels=256,
+            roi_feat_size=7,
+            reg_feat_up_ratio=2,
+            reg_pre_kernel=3,
+            reg_post_kernel=3,
+            reg_pre_num=2,
+            reg_post_num=1,
+            cls_out_channels=1024,
+            reg_offset_out_channels=256,
+            reg_cls_out_channels=256,
+            num_cls_fcs=1,
+            num_reg_fcs=0,
+            reg_class_agnostic=True,
+            norm_cfg=None,
+            bbox_coder=dict(
+                type='BucketingBBoxCoder', num_buckets=14, scale_factor=1.5),
+            loss_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
+            loss_bbox_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+            loss_bbox_reg=dict(type='SmoothL1Loss', beta=0.1,
+                               loss_weight=1.0)),
+        dict(
+            type='SABLHead',
+            num_classes=80,
+            cls_in_channels=256,
+            reg_in_channels=256,
+            roi_feat_size=7,
+            reg_feat_up_ratio=2,
+            reg_pre_kernel=3,
+            reg_post_kernel=3,
+            reg_pre_num=2,
+            reg_post_num=1,
+            cls_out_channels=1024,
+            reg_offset_out_channels=256,
+            reg_cls_out_channels=256,
+            num_cls_fcs=1,
+            num_reg_fcs=0,
+            reg_class_agnostic=True,
+            norm_cfg=None,
+            bbox_coder=dict(
+                type='BucketingBBoxCoder', num_buckets=14, scale_factor=1.3),
+            loss_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
+            loss_bbox_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+            loss_bbox_reg=dict(type='SmoothL1Loss', beta=0.1, loss_weight=1.0))
+    ]))
diff --git a/configs/sabl/sabl_cascade_rcnn_r50_fpn_1x_coco.py b/configs/sabl/sabl_cascade_rcnn_r50_fpn_1x_coco.py
@@ -0,0 +1,86 @@
+_base_ = [
+    '../_base_/models/cascade_rcnn_r50_fpn.py',
+    '../_base_/datasets/coco_detection.py',
+    '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py'
+]
+# model settings
+model = dict(
+    roi_head=dict(bbox_head=[
+        dict(
+            type='SABLHead',
+            num_classes=80,
+            cls_in_channels=256,
+            reg_in_channels=256,
+            roi_feat_size=7,
+            reg_feat_up_ratio=2,
+            reg_pre_kernel=3,
+            reg_post_kernel=3,
+            reg_pre_num=2,
+            reg_post_num=1,
+            cls_out_channels=1024,
+            reg_offset_out_channels=256,
+            reg_cls_out_channels=256,
+            num_cls_fcs=1,
+            num_reg_fcs=0,
+            reg_class_agnostic=True,
+            norm_cfg=None,
+            bbox_coder=dict(
+                type='BucketingBBoxCoder', num_buckets=14, scale_factor=1.7),
+            loss_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
+            loss_bbox_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+            loss_bbox_reg=dict(type='SmoothL1Loss', beta=0.1,
+                               loss_weight=1.0)),
+        dict(
+            type='SABLHead',
+            num_classes=80,
+            cls_in_channels=256,
+            reg_in_channels=256,
+            roi_feat_size=7,
+            reg_feat_up_ratio=2,
+            reg_pre_kernel=3,
+            reg_post_kernel=3,
+            reg_pre_num=2,
+            reg_post_num=1,
+            cls_out_channels=1024,
+            reg_offset_out_channels=256,
+            reg_cls_out_channels=256,
+            num_cls_fcs=1,
+            num_reg_fcs=0,
+            reg_class_agnostic=True,
+            norm_cfg=None,
+            bbox_coder=dict(
+                type='BucketingBBoxCoder', num_buckets=14, scale_factor=1.5),
+            loss_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
+            loss_bbox_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+            loss_bbox_reg=dict(type='SmoothL1Loss', beta=0.1,
+                               loss_weight=1.0)),
+        dict(
+            type='SABLHead',
+            num_classes=80,
+            cls_in_channels=256,
+            reg_in_channels=256,
+            roi_feat_size=7,
+            reg_feat_up_ratio=2,
+            reg_pre_kernel=3,
+            reg_post_kernel=3,
+            reg_pre_num=2,
+            reg_post_num=1,
+            cls_out_channels=1024,
+            reg_offset_out_channels=256,
+            reg_cls_out_channels=256,
+            num_cls_fcs=1,
+            num_reg_fcs=0,
+            reg_class_agnostic=True,
+            norm_cfg=None,
+            bbox_coder=dict(
+                type='BucketingBBoxCoder', num_buckets=14, scale_factor=1.3),
+            loss_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
+            loss_bbox_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+            loss_bbox_reg=dict(type='SmoothL1Loss', beta=0.1, loss_weight=1.0))
+    ]))
diff --git a/configs/sabl/sabl_faster_rcnn_r101_fpn_1x_coco.py b/configs/sabl/sabl_faster_rcnn_r101_fpn_1x_coco.py
@@ -0,0 +1,36 @@
+_base_ = [
+    '../_base_/models/faster_rcnn_r50_fpn.py',
+    '../_base_/datasets/coco_detection.py',
+    '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py'
+]
+model = dict(
+    pretrained='torchvision://resnet101',
+    backbone=dict(depth=101),
+    roi_head=dict(
+        bbox_head=dict(
+            _delete_=True,
+            type='SABLHead',
+            num_classes=80,
+            cls_in_channels=256,
+            reg_in_channels=256,
+            roi_feat_size=7,
+            reg_feat_up_ratio=2,
+            reg_pre_kernel=3,
+            reg_post_kernel=3,
+            reg_pre_num=2,
+            reg_post_num=1,
+            cls_out_channels=1024,
+            reg_offset_out_channels=256,
+            reg_cls_out_channels=256,
+            num_cls_fcs=1,
+            num_reg_fcs=0,
+            reg_class_agnostic=True,
+            norm_cfg=None,
+            bbox_coder=dict(
+                type='BucketingBBoxCoder', num_buckets=14, scale_factor=1.7),
+            loss_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
+            loss_bbox_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+            loss_bbox_reg=dict(type='SmoothL1Loss', beta=0.1,
+                               loss_weight=1.0))))
diff --git a/configs/sabl/sabl_faster_rcnn_r50_fpn_1x_coco.py b/configs/sabl/sabl_faster_rcnn_r50_fpn_1x_coco.py
@@ -0,0 +1,34 @@
+_base_ = [
+    '../_base_/models/faster_rcnn_r50_fpn.py',
+    '../_base_/datasets/coco_detection.py',
+    '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py'
+]
+model = dict(
+    roi_head=dict(
+        bbox_head=dict(
+            _delete_=True,
+            type='SABLHead',
+            num_classes=80,
+            cls_in_channels=256,
+            reg_in_channels=256,
+            roi_feat_size=7,
+            reg_feat_up_ratio=2,
+            reg_pre_kernel=3,
+            reg_post_kernel=3,
+            reg_pre_num=2,
+            reg_post_num=1,
+            cls_out_channels=1024,
+            reg_offset_out_channels=256,
+            reg_cls_out_channels=256,
+            num_cls_fcs=1,
+            num_reg_fcs=0,
+            reg_class_agnostic=True,
+            norm_cfg=None,
+            bbox_coder=dict(
+                type='BucketingBBoxCoder', num_buckets=14, scale_factor=1.7),
+            loss_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
+            loss_bbox_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+            loss_bbox_reg=dict(type='SmoothL1Loss', beta=0.1,
+                               loss_weight=1.0))))
diff --git a/configs/sabl/sabl_retinanet_r101_fpn_1x_coco.py b/configs/sabl/sabl_retinanet_r101_fpn_1x_coco.py
@@ -0,0 +1,52 @@
+_base_ = [
+    '../_base_/models/retinanet_r50_fpn.py',
+    '../_base_/datasets/coco_detection.py',
+    '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py'
+]
+# model settings
+model = dict(
+    pretrained='torchvision://resnet101',
+    backbone=dict(depth=101),
+    bbox_head=dict(
+        _delete_=True,
+        type='SABLRetinaHead',
+        num_classes=80,
+        in_channels=256,
+        stacked_convs=4,
+        feat_channels=256,
+        approx_anchor_generator=dict(
+            type='AnchorGenerator',
+            octave_base_scale=4,
+            scales_per_octave=3,
+            ratios=[0.5, 1.0, 2.0],
+            strides=[8, 16, 32, 64, 128]),
+        square_anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales=[4],
+            strides=[8, 16, 32, 64, 128]),
+        bbox_coder=dict(
+            type='BucketingBBoxCoder', num_buckets=14, scale_factor=3.0),
+        loss_cls=dict(
+            type='FocalLoss',
+            use_sigmoid=True,
+            gamma=2.0,
+            alpha=0.25,
+            loss_weight=1.0),
+        loss_bbox_cls=dict(
+            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.5),
+        loss_bbox_reg=dict(
+            type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.5)))
+# training and testing settings
+train_cfg = dict(
+    assigner=dict(
+        type='ApproxMaxIoUAssigner',
+        pos_iou_thr=0.5,
+        neg_iou_thr=0.4,
+        min_pos_iou=0.0,
+        ignore_iof_thr=-1),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+# optimizer
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)