\nchange public models with my yolo folds and get LB 0.441
\n
![\"\"](\"https://storage.yandexcloud.net/kaggle/lung.png\")
\nBlue -- my mmdet, Green -- public cascade rcnn,
Strange thing in my detection experiments
\nI have push my mmdetection model into
\nhttps://www.kaggle.com/sreevishnudamodaran/siim-effnetv2-l-cascadercnn-mmdetection-infer
\nI have trained mmdet on own preprocessing and label extraction and have 1 class detection CV ~0.58
\nI change preprocessing to save image with keeping aspect ratio and just scale it with 0.4.
\nbase LB -- 0.612
\n my mmdet -- 0.527
same story with yolo
\nchange public models with my yolo folds and get LB 0.441
\n
\nBlue -- my mmdet, Green -- public cascade rcnn,
Strange thing in my detection experiments
\nI have push my mmdetection model into
\nhttps://www.kaggle.com/sreevishnudamodaran/siim-effnetv2-l-cascadercnn-mmdetection-infer
\nI have trained mmdet on own preprocessing and label extraction and have 1 class detection CV ~0.58
\nI change preprocessing to save image with keeping aspect ratio and just scale it with 0.4.
\nbase LB -- 0.612
\n my mmdet -- 0.527
same story with yolo
\nchange public models with my yolo folds and get LB 0.441
\n
\nBlue -- my mmdet, Green -- public cascade rcnn,
The drop in LB score (between base and your mmdet) is 0.085 which is 0.510 mAP on opacity detection. Basically detection is changing from 0.55 mAP to 0.04 mAP. In other words, your detection mAP is basically zero.
\nWhen i look at your picture of green and blue boxes, it appears that your scaling is wrong. The green bboxes are not on the lungs. I would double check that you are scaling the bboxes correctly to match the test image sizes.
", "rawMarkdown": "The drop in LB score (between base and your mmdet) is 0.085 which is 0.510 mAP on opacity detection. Basically detection is changing from 0.55 mAP to 0.04 mAP. In other words, your detection mAP is basically zero.\n\nWhen i look at your picture of green and blue boxes, it appears that your scaling is wrong. The green bboxes are not on the lungs. I would double check that you are scaling the bboxes correctly to match the test image sizes.", "votes": null }, { "id": "1451467", "postDate": "08/05/2021 10:56:08", "content": "Thank you for answer
\nI saw this scale problem. But it in public results with my visualization code. I double-check and not found scaling error. Big chance that my postprocessing algorithm is wrong and i cant find it.
What are you using to make inference with mmdet? If you are using the inference_detector function, you do not need to make any changes to the outputs - they are already in x1, y1, x2, y2 format. So, if you pass an image of an un-resized xray to the function, you should get the coordinates in the required format.
", "rawMarkdown": "What are you using to make inference with mmdet? If you are using the inference_detector function, you do not need to make any changes to the outputs - they are already in x1, y1, x2, y2 format. So, if you pass an image of an un-resized xray to the function, you should get the coordinates in the required format.", "votes": null }, { "id": "1452126", "postDate": "08/05/2021 14:23:37", "content": "@novice03
\nIn most of public notebooks detection runs on presaved 512x512 images and then apply img_W,Img_H transform
\nIn my case used presaved images with 0.4 scale image and just dived box on 0.4 at postprocessing
\nWhen i plot images from public it misscaled on my image. If I change img_W,Img_H -> img_H,Img_W in postprocess it looks ok on my visualization but score is 0.528 (from 0.612)
The drop in LB score (between base and your mmdet) is 0.085 which is 0.510 mAP on opacity detection. Basically detection is changing from 0.55 mAP to 0.04 mAP. In other words, your detection mAP is basically zero.
\nWhen i look at your picture of green and blue boxes, it appears that your scaling is wrong. The green bboxes are not on the lungs. I would double check that you are scaling the bboxes correctly to match the test image sizes.
", "votes": null, "replies": [ { "id": 1451467, "author_name": "nicksergievskiy", "author_url": "", "post_date": "08/05/2021 10:56:08", "content": "Thank you for answer
\nI saw this scale problem. But it in public results with my visualization code. I double-check and not found scaling error. Big chance that my postprocessing algorithm is wrong and i cant find it.
What are you using to make inference with mmdet? If you are using the inference_detector function, you do not need to make any changes to the outputs - they are already in x1, y1, x2, y2 format. So, if you pass an image of an un-resized xray to the function, you should get the coordinates in the required format.
", "votes": null, "replies": [] } ] }, { "id": 1452126, "author_name": "nicksergievskiy", "author_url": "", "post_date": "08/05/2021 14:23:37", "content": "@novice03
\nIn most of public notebooks detection runs on presaved 512x512 images and then apply img_W,Img_H transform
\nIn my case used presaved images with 0.4 scale image and just dived box on 0.4 at postprocessing
\nWhen i plot images from public it misscaled on my image. If I change img_W,Img_H -> img_H,Img_W in postprocess it looks ok on my visualization but score is 0.528 (from 0.612)