I am playing around with an oriented RCNN with a custom satellite DOTA dataset in MMRotate. When I train the Oriented RCNN, the bounding box regression loss starts very low (loss ~ 0.009) and almost consistently increases when the RPN regression loss decreases! Some facts about the dataset:
- around 60.000 images.
- each images can contain between 0 - 4 objects.
- the images are large (1024x1024) and the objects tend to be very small.
I have looked at my data and angle definitions and am convinced that these are set correctly.
Does anyone have experienced anything similar?
Here is the config for mmrotate oriented rcnn:
_base_ = ['../_base_/datasets/ssdd.py', '../_base_/schedules/schedule_3x.py','../_base_/default_runtime.py']angle_version = 'le90'model = dict( type='OrientedRCNN', backbone=dict( type='ResNet', depth=50, num_stages=4, out_indices=(0, 1, 2, 3), frozen_stages=-1, norm_cfg=dict(type='BN', requires_grad=True), norm_eval=True, style='pytorch', init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet50')), neck=dict( type='FPN', in_channels=[256, 512, 1024, 2048], out_channels=256, num_outs=5), rpn_head=dict( type='OrientedRPNHead', in_channels=256, feat_channels=256, version=angle_version, anchor_generator=dict( type='AnchorGenerator', scales=[8], ratios=[0.5, 1.0, 2.0], strides=[4, 8, 16, 32, 64]), bbox_coder=dict( type='MidpointOffsetCoder', angle_range=angle_version, target_means=[0.0, 0.0, 0.0, 0.0, 0.0, 0.0], target_stds=[1.0, 1.0, 1.0, 1.0, 0.5, 0.5]), loss_cls=dict( type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0), loss_bbox=dict( type='SmoothL1Loss', beta=0.1111111111111111, loss_weight=1.0)), roi_head=dict( type='OrientedStandardRoIHead', bbox_roi_extractor=dict( type='RotatedSingleRoIExtractor', roi_layer=dict( type='RoIAlignRotated', out_size=7, sample_num=2, clockwise=True), out_channels=256, featmap_strides=[4, 8, 16, 32]), bbox_head=dict( type='RotatedShared2FCBBoxHead', in_channels=256, fc_out_channels=1024, roi_feat_size=7, num_classes=1, bbox_coder=dict( type='DeltaXYWHAOBBoxCoder', angle_range=angle_version, norm_factor=None, edge_swap=True, proj_xy=True, target_means=(.0, .0, .0, .0, .0), target_stds=(0.1, 0.1, 0.2, 0.2, 0.1)), reg_class_agnostic=True, loss_cls=dict( type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0), loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0))), train_cfg=dict( rpn=dict( assigner=dict( type='MaxIoUAssigner', pos_iou_thr=0.7, neg_iou_thr=0.3, min_pos_iou=0.3, match_low_quality=True, ignore_iof_thr=-1), sampler=dict( type='RandomSampler', num=256, pos_fraction=0.5, neg_pos_ub=-1, add_gt_as_proposals=False), allowed_border=0, pos_weight=-1, debug=False), rpn_proposal=dict( nms_pre=2000, max_per_img=2000, nms=dict(type='nms', iou_threshold=0.8), min_bbox_size=0), rcnn=dict( assigner=dict( type='MaxIoUAssigner', pos_iou_thr=0.5, neg_iou_thr=0.5, min_pos_iou=0.5, match_low_quality=False, iou_calculator=dict(type='RBboxOverlaps2D'), ignore_iof_thr=-1), sampler=dict( type='RRandomSampler', num=512, pos_fraction=0.25, neg_pos_ub=-1, add_gt_as_proposals=True), pos_weight=-1, debug=False)), test_cfg=dict( rpn=dict( nms_pre=2000, max_per_img=2000, nms=dict(type='nms', iou_threshold=0.8), min_bbox_size=0), rcnn=dict( nms_pre=2000, min_bbox_size=0, score_thr=0.05, nms=dict(iou_thr=0.1), max_per_img=2000)))img_norm_cfg = dict( mean=[21.55, 21.55, 21.55], std=[24.42, 24.42, 24.42], to_rgb=True)train_pipeline = [ dict(type='LoadImageFromFile'), dict(type='LoadAnnotations', with_bbox=True), dict(type='RResize', img_scale=(608, 608)), dict( type='RRandomFlip', flip_ratio=[0.25, 0.25, 0.25], direction=['horizontal', 'vertical', 'diagonal'], version=angle_version), dict(type='Normalize', **img_norm_cfg), dict(type='Pad', size_divisor=32), dict(type='DefaultFormatBundle'), dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])]data = dict( train=dict(pipeline=train_pipeline, version=angle_version), val=dict(version=angle_version), test=dict(version=angle_version))optimizer = dict(lr=0.005)# evaluationevaluation = dict(interval=72, metric='mAP')# learning policylr_config = dict( policy='step', warmup='linear', warmup_iters=500, warmup_ratio=1.0 / 3, step=[65, 71])runner = dict(type='EpochBasedRunner', max_epochs=72)checkpoint_config = dict(interval=12)