There is 1 special point in this competition: there are many image annotated by multiple radiologists independently. Many boxes are of very different sizes, even though they mark the same object. Maybe, You should be careful when using NMS (non maximum suppression). Suppose that in an image, there are two GROUND TRUTH bboxes overlap with IOU > 0.9. BOTH of these boxes are true labels (GROUND TRUTH) because they are independently annotated by 2 radiologists. Your model return 2 output bboxes corresponding to labels. In this case, if you use NMS you will remove 1 box in output even though it is correct result.
\nAny ideal ?
", "messages": [ { "id": 1134216, "postDate": "2021-01-01T02:25:33.860Z", "content": "There is 1 special point in this competition: there are many image annotated by multiple radiologists independently. Many boxes are of very different sizes, even though they mark the same object. Maybe, You should be careful when using NMS (non maximum suppression). Suppose that in an image, there are two GROUND TRUTH bboxes overlap with IOU > 0.9. BOTH of these boxes are true labels (GROUND TRUTH) because they are independently annotated by 2 radiologists. Your model return 2 output bboxes corresponding to labels. In this case, if you use NMS you will remove 1 box in output even though it is correct result.
\nAny ideal ?
", "rawMarkdown": "There is 1 special point in this competition: there are many image annotated by multiple radiologists independently. Many boxes are of very different sizes, even though they mark the same object. Maybe, You should be careful when using NMS (non maximum suppression). Suppose that in an image, there are two GROUND TRUTH bboxes overlap with IOU > 0.9. BOTH of these boxes are true labels (GROUND TRUTH) because they are independently annotated by 2 radiologists. Your model return 2 output bboxes corresponding to labels. In this case, if you use NMS you will remove 1 box in output even though it is correct result.\n\nAny ideal ?", "votes": 6 }, { "id": 1138242, "postDate": "2021-01-04T14:21:08.687Z", "content": "I'm currently taking the average of (x1, y1, x2, y2) if these boxes are overlapping with iou smaller than 0.4
", "rawMarkdown": "I'm currently taking the average of (x1, y1, x2, y2) if these boxes are overlapping with iou smaller than 0.4" }, { "id": 1135539, "postDate": "2021-01-02T10:30:27.710Z", "content": "I think simply, overlapping of all radiologists is consensus 😄
", "rawMarkdown": "I think simply, overlapping of all radiologists is **consensus** 😄" }, { "id": 1134408, "postDate": "2021-01-01T09:17:37.133Z", "content": "I had the idea that we can use NMS for boxes with IoU above a certain threshold and then run Object detection algorithms on the images
", "rawMarkdown": "I had the idea that we can use NMS for boxes with IoU above a certain threshold and then run Object detection algorithms on the images", "replies": [ { "id": 1134464, "postDate": "2021-01-01T10:12:36.450Z", "content": "Suppose that in an image, there are two GROUND TRUTH bboxes overlap with IOU > 0.9. BOTH of these boxes are true labels (GROUND TRUTH) because they are independently annotated by 2 radiologists. Your model return 2 output bboxes corresponding to labels. In this case, if you use NMS you will remove 1 box in output even though it is correct result.
", "rawMarkdown": "Suppose that in an image, there are two GROUND TRUTH bboxes overlap with IOU > 0.9. BOTH of these boxes are true labels (GROUND TRUTH) because they are independently annotated by 2 radiologists. Your model return 2 output bboxes corresponding to labels. In this case, if you use NMS you will remove 1 box in output even though it is correct result.\n" }, { "id": 1134710, "postDate": "2021-01-01T14:08:08.867Z", "content": "Check this discussion
\nThere are no overlapping bounding boxes in the test set
\nhttps://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/207969#1134645
many thanks,
", "rawMarkdown": "many thanks, " } ] } ], "comments": [ { "id": 1138242, "author_name": "Phat Tran", "author_url": "", "post_date": "2021-01-04T14:21:08.687000", "content": "I'm currently taking the average of (x1, y1, x2, y2) if these boxes are overlapping with iou smaller than 0.4
", "votes": 0, "replies": [] }, { "id": 1135539, "author_name": "tamnv", "author_url": "", "post_date": "2021-01-02T10:30:27.710000", "content": "I think simply, overlapping of all radiologists is consensus 😄
", "votes": 0, "replies": [] }, { "id": 1134408, "author_name": "Debarshi Chanda", "author_url": "", "post_date": "2021-01-01T09:17:37.133000", "content": "I had the idea that we can use NMS for boxes with IoU above a certain threshold and then run Object detection algorithms on the images
", "votes": 0, "replies": [ { "id": 1134464, "author_name": "Trung Thanh Nguyen ", "author_url": "", "post_date": "2021-01-01T10:12:36.450000", "content": "Suppose that in an image, there are two GROUND TRUTH bboxes overlap with IOU > 0.9. BOTH of these boxes are true labels (GROUND TRUTH) because they are independently annotated by 2 radiologists. Your model return 2 output bboxes corresponding to labels. In this case, if you use NMS you will remove 1 box in output even though it is correct result.
", "votes": 0, "replies": [] }, { "id": 1134710, "author_name": "Debarshi Chanda", "author_url": "", "post_date": "2021-01-01T14:08:08.867000", "content": "Check this discussion
\nThere are no overlapping bounding boxes in the test set
\nhttps://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/207969#1134645
many thanks,
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1134216": "There is 1 special point in this competition: there are many image annotated by multiple radiologists independently. Many boxes are of very different sizes, even though they mark the same object. Maybe, You should be careful when using NMS (non maximum suppression). Suppose that in an image, there are two GROUND TRUTH bboxes overlap with IOU > 0.9. BOTH of these boxes are true labels (GROUND TRUTH) because they are independently annotated by 2 radiologists. Your model return 2 output bboxes corresponding to labels. In this case, if you use NMS you will remove 1 box in output even though it is correct result.\n\nAny ideal ?", "1138242": "I'm currently taking the average of (x1, y1, x2, y2) if these boxes are overlapping with iou smaller than 0.4", "1135539": "I think simply, overlapping of all radiologists is **consensus** 😄", "1134408": "I had the idea that we can use NMS for boxes with IoU above a certain threshold and then run Object detection algorithms on the images" } }