本次比赛的目标是通过建立一个在珊瑚礁水下视频上训练的物体检测模型，实时准确地识别海星。
比赛方案：工具库：mmdetection,yolov5
开源代码链接：the framework of ensemble | Kaggle
比赛最终的方案是fasterrcnn和yolov5通过wbf模型融合
定义模型
config_file = f'../input/fastrcnn2000/faster_rcnn_cbv2d1_r50_fpn_1x_coco.py'checkpoint_file = '../input/faster-rcnn/faster_rcnn_cbv2/epoch_8.pth' m1 = init_detector(config_file, checkpoint_file, device='cuda:0')m2 = torch.hub.load('../input/yolov5-lib-ds', 'custom', path='../input/yolov5s6/f2_sub2.pt', source='local', force_reload=True) # local repom2.conf = 0.48
wbf示例代码
def wbf_one_img_result(result_cbnet, result_yolo, IOU_WBF=0.2): boxes_list = [[],[]] scores_list = [[], []] labels_list = [[], []] weights = [1, 1] iou_thr = 0.5 skip_box_thr = 0.3 #sigma = 0.1 for cbnet in result_cbnet[0]: x1, y1, x2, y2, conf = cbnet boxes_list[0].append([x1/1280, y1/720, x2/1280, y2/720]) scores_list[0].append(conf) labels_list[0].append(0)for idx, row in result_yolo.pandas().xyxy[0].iterrows(): boxes_list[1].append([row.xmin/1280, row.ymin/720, row.xmax/1280, row.ymax/720]) scores_list[1].append(row.confidence) labels_list[1].append(0) boxes, scores, labels = weighted_boxes_fusion(boxes_list, scores_list, labels_list, weights=weights, iou_thr=iou_thr, skip_box_thr=skip_box_thr) length = len(scores) res = '' for idx in range(length): conf = scores[idx] xmin, ymin, xmax, ymax = boxes[idx] if conf > 0.4: res += f'{conf} {xmin*1280} {ymin*720} {(xmax-xmin)*1280} {(ymax-ymin)*720} ' return res.strip(' ')
上分技巧：
1.加大分辨率：提高yolo输入图片的size可以提高模型的性能
2.工具库：ensemble-boxes
比赛总结：最终获得前2%的成绩