Hi, I was using metric learning with/without bounding box. The bounding box is extracted by using \n https://www.kaggle.com/martinpiotte/bounding-box-model , but it seems that adding bounding box gives worse results.
\n\nI evaluate martinpiotte's bounding box model by calculating IoU based on the annotatation provided by https://www.kaggle.com/c/humpback-whale-identification/discussion/76281 , the iou is about 0.88, I think the detection result is OK, but adding bounding box to metric learning actually gives bad results, which surprised me.
\n\nMay I ask did you encounter similar issue? How did you solve it?
", "messages": [ { "id": "463080", "postDate": "01/29/2019 11:17:38", "content": "Hi, I was using metric learning with/without bounding box. The bounding box is extracted by using \n https://www.kaggle.com/martinpiotte/bounding-box-model , but it seems that adding bounding box gives worse results.
\n\nI evaluate martinpiotte's bounding box model by calculating IoU based on the annotatation provided by https://www.kaggle.com/c/humpback-whale-identification/discussion/76281 , the iou is about 0.88, I think the detection result is OK, but adding bounding box to metric learning actually gives bad results, which surprised me.
\n\nMay I ask did you encounter similar issue? How did you solve it?
", "rawMarkdown": "Hi, I was using metric learning with/without bounding box. The bounding box is extracted by using \n https://www.kaggle.com/martinpiotte/bounding-box-model , but it seems that adding bounding box gives worse results. \n\nI evaluate martinpiotte's bounding box model by calculating IoU based on the annotatation provided by https://www.kaggle.com/c/humpback-whale-identification/discussion/76281 , the iou is about 0.88, I think the detection result is OK, but adding bounding box to metric learning actually gives bad results, which surprised me.\n\n\nMay I ask did you encounter similar issue? How did you solve it?", "votes": null }, { "id": "463310", "postDate": "01/29/2019 19:32:48", "content": "I did not try without bounding boxes, all my tests are with them... but mainly I tried submissions for now without new whales on the csv... To comapre results could you ask me -> https://www.kaggle.com/c/humpback-whale-identification/discussion/77489 When I have results without bounding boxes I will post them!
", "rawMarkdown": "I did not try without bounding boxes, all my tests are with them... but mainly I tried submissions for now without new whales on the csv... To comapre results could you ask me -> https://www.kaggle.com/c/humpback-whale-identification/discussion/77489 When I have results without bounding boxes I will post them!", "votes": null }, { "id": "464105", "postDate": "01/31/2019 07:49:55", "content": "bounding box works for me.
", "rawMarkdown": "bounding box works for me.", "votes": null }, { "id": "464131", "postDate": "01/31/2019 08:36:27", "content": "Can I ask you whats your LB for your submission without new whales?
", "rawMarkdown": "Can I ask you whats your LB for your submission without new whales?", "votes": null }, { "id": "464197", "postDate": "01/31/2019 10:59:47", "content": "May I ask how do you get bounding box? Do you use martinpiotte's pretrained model from https://www.kaggle.com/martinpiotte/bounding-box-model ?
", "rawMarkdown": "May I ask how do you get bounding box? Do you use martinpiotte's pretrained model from https://www.kaggle.com/martinpiotte/bounding-box-model ?", "votes": null }, { "id": "464567", "postDate": "02/01/2019 04:56:12", "content": "Actually I don't know how to use this code, so I trained one myself.
", "rawMarkdown": "Actually I don't know how to use this code, so I trained one myself.", "votes": null }, { "id": "464568", "postDate": "02/01/2019 05:00:49", "content": "I didn't try Without new whale. I think the most important task in this competition is dealing with the 'new whale' .
", "rawMarkdown": "I didn't try Without new whale. I think the most important task in this competition is dealing with the 'new whale' .", "votes": null }, { "id": "470523", "postDate": "02/13/2019 05:43:58", "content": "Same observation here. Just report that the bounding boxes has tons of low quality detections, which is a disaster to the metric learning model proposed by Piotte. By hard example mining, the model would easily overfit to the noisy data, which may explain the worse result.
\n\nMy solution is letting hard example mining work for several epochs to overfit, then remove all the training samples it overfits to. Kinda works
", "rawMarkdown": "Same observation here. Just report that the bounding boxes has tons of low quality detections, which is a disaster to the metric learning model proposed by Piotte. By hard example mining, the model would easily overfit to the noisy data, which may explain the worse result.\n\nMy solution is letting hard example mining work for several epochs to overfit, then remove all the training samples it overfits to. Kinda works", "votes": null }, { "id": "470643", "postDate": "02/13/2019 09:55:46", "content": "Simple thing to check bad bboxes: file size: if too small, then it is probably bad. Same with std(image). Then you can manually check candidates.\nFor me, this thing + public bboxes were enough.
", "rawMarkdown": "Simple thing to check bad bboxes: file size: if too small, then it is probably bad. Same with std(image). Then you can manually check candidates.\nFor me, this thing + public bboxes were enough.", "votes": null } ], "comments": [ { "id": 463310, "author_name": "maparla", "author_url": "", "post_date": "01/29/2019 19:32:48", "content": "I did not try without bounding boxes, all my tests are with them... but mainly I tried submissions for now without new whales on the csv... To comapre results could you ask me -> https://www.kaggle.com/c/humpback-whale-identification/discussion/77489 When I have results without bounding boxes I will post them!
", "votes": null, "replies": [] }, { "id": 464105, "author_name": "xf1994", "author_url": "", "post_date": "01/31/2019 07:49:55", "content": "bounding box works for me.
", "votes": null, "replies": [ { "id": 464131, "author_name": "maparla", "author_url": "", "post_date": "01/31/2019 08:36:27", "content": "Can I ask you whats your LB for your submission without new whales?
", "votes": null, "replies": [] }, { "id": 464197, "author_name": "zjucor", "author_url": "", "post_date": "01/31/2019 10:59:47", "content": "May I ask how do you get bounding box? Do you use martinpiotte's pretrained model from https://www.kaggle.com/martinpiotte/bounding-box-model ?
", "votes": null, "replies": [] }, { "id": 464567, "author_name": "xf1994", "author_url": "", "post_date": "02/01/2019 04:56:12", "content": "Actually I don't know how to use this code, so I trained one myself.
", "votes": null, "replies": [] }, { "id": 464568, "author_name": "xf1994", "author_url": "", "post_date": "02/01/2019 05:00:49", "content": "I didn't try Without new whale. I think the most important task in this competition is dealing with the 'new whale' .
", "votes": null, "replies": [] } ] }, { "id": 470523, "author_name": "atm584", "author_url": "", "post_date": "02/13/2019 05:43:58", "content": "Same observation here. Just report that the bounding boxes has tons of low quality detections, which is a disaster to the metric learning model proposed by Piotte. By hard example mining, the model would easily overfit to the noisy data, which may explain the worse result.
\n\nMy solution is letting hard example mining work for several epochs to overfit, then remove all the training samples it overfits to. Kinda works
", "votes": null, "replies": [] }, { "id": 470643, "author_name": "oldufo", "author_url": "", "post_date": "02/13/2019 09:55:46", "content": "Simple thing to check bad bboxes: file size: if too small, then it is probably bad. Same with std(image). Then you can manually check candidates.\nFor me, this thing + public bboxes were enough.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "463080": "Hi, I was using metric learning with/without bounding box. The bounding box is extracted by using \n https://www.kaggle.com/martinpiotte/bounding-box-model , but it seems that adding bounding box gives worse results. \n\nI evaluate martinpiotte's bounding box model by calculating IoU based on the annotatation provided by https://www.kaggle.com/c/humpback-whale-identification/discussion/76281 , the iou is about 0.88, I think the detection result is OK, but adding bounding box to metric learning actually gives bad results, which surprised me.\n\n\nMay I ask did you encounter similar issue? How did you solve it?", "463310": "I did not try without bounding boxes, all my tests are with them... but mainly I tried submissions for now without new whales on the csv... To comapre results could you ask me -> https://www.kaggle.com/c/humpback-whale-identification/discussion/77489 When I have results without bounding boxes I will post them!", "464105": "bounding box works for me.", "464131": "Can I ask you whats your LB for your submission without new whales?", "464197": "May I ask how do you get bounding box? Do you use martinpiotte's pretrained model from https://www.kaggle.com/martinpiotte/bounding-box-model ?", "464567": "Actually I don't know how to use this code, so I trained one myself.", "464568": "I didn't try Without new whale. I think the most important task in this competition is dealing with the 'new whale' .", "470523": "Same observation here. Just report that the bounding boxes has tons of low quality detections, which is a disaster to the metric learning model proposed by Piotte. By hard example mining, the model would easily overfit to the noisy data, which may explain the worse result.\n\nMy solution is letting hard example mining work for several epochs to overfit, then remove all the training samples it overfits to. Kinda works", "470643": "Simple thing to check bad bboxes: file size: if too small, then it is probably bad. Same with std(image). Then you can manually check candidates.\nFor me, this thing + public bboxes were enough." }, "source": "meta" }