6/xx Hubmap 2023 10th place solutions by sugupoko Training flow Pretrain Yolov7seg Dataset : ds2 only1fold 、100epoch Yolov7seg Predictions ds2&ds3 Pseudo Label ds2&ds3 Train Yolov7seg Dataset : ds1 5fold、100epoch Pretrain Yolov7seg Dataset: pseudo ds2&ds3 only1fold 、100epoch weight Train Yolov7seg Dataset : ds1 5fold 、100epoch Inference flow weight YoloV7 Fold0,1,4 Post Processing Get Largest mask with connected component From @fnands notebook A quick YOLOv7 Baseline | Kaggle Several parameters are changed. Mainly resolution(512=>640) From @fnands notebook A Quick YOLOv7 Baseline [Inference] | Kaggle Revise the NMS code. Make it utilize the mask of the fold selected by NMS. And, add post processing. Only the largest masks were retained. Resolution 640 Conf_th = 0.001 Iou_th = 0.55 LB0.359、PB:0.418 LB0.517、PB:0.549 weight Pickup data w/ predicted conf Conf at instance >0.5 Average conf per tile > 0.6 Sample Result Over detect!!!! Remove dilate 全体像
7/xx Hubmap 2023 10th place solutions by sugupoko Training flow Pretrain Yolov7seg Dataset : ds2 only1fold 、100epoch Yolov7seg Predictions ds2&ds3 Pseudo Label ds2&ds3 Train Yolov7seg Dataset : ds1 5fold、100epoch Pretrain Yolov7seg Dataset: pseudo ds2&ds3 only1fold 、100epoch weight Train Yolov7seg Dataset : ds1 5fold 、100epoch Inference flow weight YoloV7 Fold0,1,4 Post Processing Get Largest mask with connected component From @fnands notebook A quick YOLOv7 Baseline | Kaggle Several parameters are changed. Mainly resolution(512=>640) From @fnands notebook A Quick YOLOv7 Baseline [Inference] | Kaggle Revise the NMS code. Make it utilize the mask of the fold selected by NMS. And, add post processing. Only the largest masks were retained. Resolution 640 Conf_th = 0.001 Iou_th = 0.55 LB0.359、PB:0.418 LB0.517、PB:0.549 weight Pickup data w/ predicted conf Conf at instance >0.5 Average conf per tile > 0.6 Sample Result Over detect!!!! Remove dilate 独自ポイント: ①ドメインをなるべく広げる事前学習&②専門家が付けたラベルに寄せる ②専門家が付けたラベルに寄せるの部分 Discussionにて、素人ラベラーのデータを混ぜるとセグメン テーションの精度が低下することが分かっていたので、疑似ラ ベル付与ミスのない専門家のみのデータでFinetuning ①ドメインをなるべく広げる事前学習の部分 画像のドメインが違うことはEDAで分かっていたので、未 ラベル画像すべてを使うように疑似ラベルを利用。 推論のConfidenceが高いもの&画像全体のインスタン スの平均が高い画像のみを選定
8/xx Hubmap 2023 10th place solutions by sugupoko Training flow Pretrain Yolov7seg Dataset : ds2 only1fold 、100epoch Yolov7seg Predictions ds2&ds3 Pseudo Label ds2&ds3 Train Yolov7seg Dataset : ds1 5fold、100epoch Pretrain Yolov7seg Dataset: pseudo ds2&ds3 only1fold 、100epoch weight Train Yolov7seg Dataset : ds1 5fold 、100epoch Inference flow weight YoloV7 Fold0,1,4 Post Processing Get Largest mask with connected component From @fnands notebook A quick YOLOv7 Baseline | Kaggle Several parameters are changed. Mainly resolution(512=>640) From @fnands notebook A Quick YOLOv7 Baseline [Inference] | Kaggle Revise the NMS code. Make it utilize the mask of the fold selected by NMS. And, add post processing. Only the largest masks were retained. Resolution 640 Conf_th = 0.001 Iou_th = 0.55 LB0.359、PB:0.418 LB0.517、PB:0.549 weight Pickup data w/ predicted conf Conf at instance >0.5 Average conf per tile > 0.6 Sample Result Over detect!!!! Remove dilate コンペに合わせた調整:①Yoloの入力解像度を変更②推論時のパラメータを変更 ①解像度の変更 YoloのSegmentation headが入力解像度に対して1/4になるので、 入力をなるべく大きくするように変更。 配布データは512x512だが、640x640に上げて学習。 事前学習モデルがあるので、ネットワークの変更はできなかった。。。 ②推論時パラメータの変更 LBを見ながら、閾値を調整。 未知のデータに対してはConfが低く、過検出の方が良い結果になる コトが分かる。