top 2% solution
Detector: Faster RCNN
backbone: ResNext101-32d-4
Detector: Faster RCNN
backbone: ResNet50
etc: attention
Detector: Faster RCNN
backbone: RegNet12
Detector: DetectoRS
backbone: ResNet50
Detector: DetectoRS
backbone: ResNext101-32d-4
ratio: [0.5, 1.0, 2.0] -> [0.33, 0.5, 1.0, 2.0, 3.0]
scale: [8] -> [1, 2, 4, 8, 16]
Detector: VFNet, Cascade RCNN, Cascade mask RCNN, Hybrid Task Casacde, Generalized Focal Loss
Backbone: EfficientNet, DenseNet
etc: pseudo labeling, segmentation, external data (NIH), Large box detector, small box detector