{ "id": 303793, "title": "0.628 @ 6000 img_size with yolov5s", "url": "/competitions/tensorflow-great-barrier-reef/discussion/303793", "author_name": "", "post_date": "2022-01-29T12:40:22.819185Z", "votes": 13, "comment_count": 10, "views": 0, "content": "

Please find the link for the model which gets 0.628 in LB with yolov5s
\nhttps://www.kaggle.com/dhakshiin1601/cots-balaji-models

\n

Use the following notebook
\nhttps://www.kaggle.com/kocha1/only-yolov5-tracking-lb-642

\n

IMG_SIZE = 6000
\nCONF = 0.10
\nIOU = 0.20

\n

I kept CONF and IOU fixed and I tried inference at multiple image size
\n0.617 - 9000
\n0.628 - 6000
\n0.562 - 4200
\n0.540 - 3600

\n

At a size of 6000, there's the possibility for us to use 3 models during the inference.
\nI hope the model and the results that I shared are useful

", "messages": [ { "id": "1668195", "postDate": "01/29/2022 12:40:22", "content": "

Please find the link for the model which gets 0.628 in LB with yolov5s
\nhttps://www.kaggle.com/dhakshiin1601/cots-balaji-models

\n

Use the following notebook
\nhttps://www.kaggle.com/kocha1/only-yolov5-tracking-lb-642

\n

IMG_SIZE = 6000
\nCONF = 0.10
\nIOU = 0.20

\n

I kept CONF and IOU fixed and I tried inference at multiple image size
\n0.617 - 9000
\n0.628 - 6000
\n0.562 - 4200
\n0.540 - 3600

\n

At a size of 6000, there's the possibility for us to use 3 models during the inference.
\nI hope the model and the results that I shared are useful

", "rawMarkdown": "Please find the link for the model which gets 0.628 in LB with yolov5s\nhttps://www.kaggle.com/dhakshiin1601/cots-balaji-models\n\nUse the following notebook\nhttps://www.kaggle.com/kocha1/only-yolov5-tracking-lb-642\n\nIMG_SIZE = 6000\nCONF = 0.10\nIOU = 0.20\n\nI kept CONF and IOU fixed and I tried inference at multiple image size \n0.617 - 9000\n0.628 - 6000\n0.562 - 4200\n0.540 - 3600\n\nAt a size of 6000, there's the possibility for us to use 3 models during the inference.\nI hope the model and the results that I shared are useful", "votes": null }, { "id": "1668215", "postDate": "01/29/2022 12:55:22", "content": "

no matter what the public lb score is,IMHO IoU 0.20 is super risky for private lb

", "rawMarkdown": "no matter what the public lb score is,IMHO IoU 0.20 is super risky for private lb", "votes": null }, { "id": "1668855", "postDate": "01/30/2022 05:46:39", "content": "

Thanks, @mobassir for the suggestion.

\n

I am getting nearly the same LB(0.628) for an IOU of 0.4.

\n

First I wanted to play around with the size, now I am changing the CONF and IOU. If I get any better performance, I will let you know 👍

", "rawMarkdown": "Thanks, @mobassir for the suggestion.\n\nI am getting nearly the same LB(0.628) for an IOU of 0.4. \n\nFirst I wanted to play around with the size, now I am changing the CONF and IOU. If I get any better performance, I will let you know 👍", "votes": null }, { "id": "1669936", "postDate": "01/31/2022 04:50:55", "content": "

I don't understand this, could you give a detailed explaination？ Thank you！

", "rawMarkdown": "I don't understand this, could you give a detailed explaination？ Thank you！", "votes": null }, { "id": "1669941", "postDate": "01/31/2022 04:58:30", "content": "

Hi, how to choose IOU thres if two results(IOU thres1=0.2,IOU thres2=0.4) are the same, higher or lower? Thanks !

", "rawMarkdown": "Hi, how to choose IOU thres if two results(IOU thres1=0.2,IOU thres2=0.4) are the same, higher or lower? Thanks !", "votes": null }, { "id": "1669963", "postDate": "01/31/2022 05:20:05", "content": "

https://medium.com/analytics-vidhya/iou-intersection-over-union-705a39e7acef

", "rawMarkdown": "https://medium.com/analytics-vidhya/iou-intersection-over-union-705a39e7acef", "votes": null }, { "id": "1670060", "postDate": "01/31/2022 07:29:26", "content": "

Okay, Thanks~

", "rawMarkdown": "Okay, Thanks~", "votes": null }, { "id": "1671011", "postDate": "02/01/2022 04:55:53", "content": "

I didn't notice much impact when I change IOU for my models. As the competition metric is evaluated from an IOU of 0.3, I think it's good to have IOU as 0.3.

", "rawMarkdown": "I didn't notice much impact when I change IOU for my models. As the competition metric is evaluated from an IOU of 0.3, I think it's good to have IOU as 0.3.", "votes": null }, { "id": "1671032", "postDate": "02/01/2022 05:23:42", "content": "

Sounds good, Thanks!

", "rawMarkdown": "Sounds good, Thanks!", "votes": null }, { "id": "1671045", "postDate": "02/01/2022 05:40:22", "content": "

I have a question. I trained yolov5s6 with size3600 and then used the following notebook https://www.kaggle.com/kocha1/only-yolov5-tracking-lb-642 for inference. conf0.34 iou0.5, in the case where I only change the img-size of inference:
\nimg-size 3600 lb0.579
\nimg-size 9000 lb0.577
\nWhy does my self-trained model infer no change at large scale?

", "rawMarkdown": "I have a question. I trained yolov5s6 with size3600 and then used the following notebook [https://www.kaggle.com/kocha1/only-yolov5-tracking-lb-642 ](url)for inference. conf0.34 iou0.5, in the case where I only change the img-size of inference:\nimg-size 3600 lb0.579\nimg-size 9000 lb0.577\nWhy does my self-trained model infer no change at large scale?", "votes": null }, { "id": "1676991", "postDate": "02/05/2022 12:44:01", "content": "

0.628 - 6000
\nIn these experiments all the other hyperparameters were the same, i mean you used the default hyperparameters? Only changing image size is giving you 0.628.

", "rawMarkdown": "`0.628 - 6000`\nIn these experiments all the other hyperparameters were the same, i mean you used the default hyperparameters? Only changing image size is giving you 0.628.", "votes": null } ], "comments": [ { "id": 1668215, "author_name": "mobassir", "author_url": "", "post_date": "01/29/2022 12:55:22", "content": "

no matter what the public lb score is,IMHO IoU 0.20 is super risky for private lb

", "votes": null, "replies": [ { "id": 1669936, "author_name": "rainfalllove", "author_url": "", "post_date": "01/31/2022 04:50:55", "content": "

I don't understand this, could you give a detailed explaination？ Thank you！

", "votes": null, "replies": [] }, { "id": 1669963, "author_name": "mobassir", "author_url": "", "post_date": "01/31/2022 05:20:05", "content": "

https://medium.com/analytics-vidhya/iou-intersection-over-union-705a39e7acef

", "votes": null, "replies": [] }, { "id": 1670060, "author_name": "rainfalllove", "author_url": "", "post_date": "01/31/2022 07:29:26", "content": "

Okay, Thanks~

", "votes": null, "replies": [] } ] }, { "id": 1668855, "author_name": "dhakshiin1601", "author_url": "", "post_date": "01/30/2022 05:46:39", "content": "

Thanks, @mobassir for the suggestion.

\n

I am getting nearly the same LB(0.628) for an IOU of 0.4.

\n

First I wanted to play around with the size, now I am changing the CONF and IOU. If I get any better performance, I will let you know 👍

", "votes": null, "replies": [ { "id": 1669941, "author_name": "rainfalllove", "author_url": "", "post_date": "01/31/2022 04:58:30", "content": "

Hi, how to choose IOU thres if two results(IOU thres1=0.2,IOU thres2=0.4) are the same, higher or lower? Thanks !

", "votes": null, "replies": [] }, { "id": 1671011, "author_name": "dhakshiin1601", "author_url": "", "post_date": "02/01/2022 04:55:53", "content": "

I didn't notice much impact when I change IOU for my models. As the competition metric is evaluated from an IOU of 0.3, I think it's good to have IOU as 0.3.

", "votes": null, "replies": [] }, { "id": 1671032, "author_name": "rainfalllove", "author_url": "", "post_date": "02/01/2022 05:23:42", "content": "

Sounds good, Thanks!

", "votes": null, "replies": [] } ] }, { "id": 1671045, "author_name": "henini", "author_url": "", "post_date": "02/01/2022 05:40:22", "content": "

I have a question. I trained yolov5s6 with size3600 and then used the following notebook https://www.kaggle.com/kocha1/only-yolov5-tracking-lb-642 for inference. conf0.34 iou0.5, in the case where I only change the img-size of inference:
\nimg-size 3600 lb0.579
\nimg-size 9000 lb0.577
\nWhy does my self-trained model infer no change at large scale?

", "votes": null, "replies": [] }, { "id": 1676991, "author_name": "soumya9977", "author_url": "", "post_date": "02/05/2022 12:44:01", "content": "

0.628 - 6000
\nIn these experiments all the other hyperparameters were the same, i mean you used the default hyperparameters? Only changing image size is giving you 0.628.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1668195": "Please find the link for the model which gets 0.628 in LB with yolov5s\nhttps://www.kaggle.com/dhakshiin1601/cots-balaji-models\n\nUse the following notebook\nhttps://www.kaggle.com/kocha1/only-yolov5-tracking-lb-642\n\nIMG_SIZE = 6000\nCONF = 0.10\nIOU = 0.20\n\nI kept CONF and IOU fixed and I tried inference at multiple image size \n0.617 - 9000\n0.628 - 6000\n0.562 - 4200\n0.540 - 3600\n\nAt a size of 6000, there's the possibility for us to use 3 models during the inference.\nI hope the model and the results that I shared are useful", "1668215": "no matter what the public lb score is,IMHO IoU 0.20 is super risky for private lb", "1668855": "Thanks, @mobassir for the suggestion.\n\nI am getting nearly the same LB(0.628) for an IOU of 0.4. \n\nFirst I wanted to play around with the size, now I am changing the CONF and IOU. If I get any better performance, I will let you know 👍", "1669936": "I don't understand this, could you give a detailed explaination？ Thank you！", "1669941": "Hi, how to choose IOU thres if two results(IOU thres1=0.2,IOU thres2=0.4) are the same, higher or lower? Thanks !", "1669963": "https://medium.com/analytics-vidhya/iou-intersection-over-union-705a39e7acef", "1670060": "Okay, Thanks~", "1671011": "I didn't notice much impact when I change IOU for my models. As the competition metric is evaluated from an IOU of 0.3, I think it's good to have IOU as 0.3.", "1671032": "Sounds good, Thanks!", "1671045": "I have a question. I trained yolov5s6 with size3600 and then used the following notebook [https://www.kaggle.com/kocha1/only-yolov5-tracking-lb-642 ](url)for inference. conf0.34 iou0.5, in the case where I only change the img-size of inference:\nimg-size 3600 lb0.579\nimg-size 9000 lb0.577\nWhy does my self-trained model infer no change at large scale?", "1676991": "`0.628 - 6000`\nIn these experiments all the other hyperparameters were the same, i mean you used the default hyperparameters? Only changing image size is giving you 0.628." }, "source": "meta" }