To evaluate the different detector solutions, I collect the open solutions on kaggle kernels.It seems that FasterRCNN has the best performance and the efficientdet has the worst performance.
\n\nTwo stage detector:
\n\n1.FasterRCNN \nimgage size:1024*1024\nmodel: resnet50\nlb:0.6687\nhttps://www.kaggle.com/pestipeti/pytorch-starter-fasterrcnn-inference
\n\nOne stage detector:
\n\n1- yolov3\nimgage size:608*608\nmodel: Darknet\nlb:0.6504\nhttps://www.kaggle.com/sumitjha19/yolov3-inference-v2-full-code
\n\n2- centernet\nimgage size:512*512\nmodel: Resnet18\nlb:0.52
\n\nhttps://www.kaggle.com/kyoshioka47/pytorch-centernet-inference-with-tta
\n\n3- yolov4\nimgage size:608*608\nmodel: Darknet\nlb:0.6184
\n\nhttps://www.kaggle.com/nvnnghia/yolov4-inference
\n\n4- efficientdet\n1) efficientdet0\nimgage size:1024*1024\nmodel: efficientnet b0\nlb:0.5322
\n\nhttps://www.kaggle.com/ateplyuk/gwd-starter-efficientdet-inference
\n\n2)efficentdet3\nimgage size:512*512\nmodel: efficientnet b3\nlb: \ncv: 0.547
\n\n5- retinanet\nimgage size:1024*1024\nmodel: resnet50\nlb: 0.5998\nhttps://www.kaggle.com/labintsevai/kerasretinanet-wheat-submission
", "messages": [ { "id": 845640, "postDate": "2020-05-13T10:37:37.407Z", "content": "To evaluate the different detector solutions, I collect the open solutions on kaggle kernels.It seems that FasterRCNN has the best performance and the efficientdet has the worst performance.
\n\nTwo stage detector:
\n\n1.FasterRCNN \nimgage size:1024*1024\nmodel: resnet50\nlb:0.6687\nhttps://www.kaggle.com/pestipeti/pytorch-starter-fasterrcnn-inference
\n\nOne stage detector:
\n\n1- yolov3\nimgage size:608*608\nmodel: Darknet\nlb:0.6504\nhttps://www.kaggle.com/sumitjha19/yolov3-inference-v2-full-code
\n\n2- centernet\nimgage size:512*512\nmodel: Resnet18\nlb:0.52
\n\nhttps://www.kaggle.com/kyoshioka47/pytorch-centernet-inference-with-tta
\n\n3- yolov4\nimgage size:608*608\nmodel: Darknet\nlb:0.6184
\n\nhttps://www.kaggle.com/nvnnghia/yolov4-inference
\n\n4- efficientdet\n1) efficientdet0\nimgage size:1024*1024\nmodel: efficientnet b0\nlb:0.5322
\n\nhttps://www.kaggle.com/ateplyuk/gwd-starter-efficientdet-inference
\n\n2)efficentdet3\nimgage size:512*512\nmodel: efficientnet b3\nlb: \ncv: 0.547
\n\n5- retinanet\nimgage size:1024*1024\nmodel: resnet50\nlb: 0.5998\nhttps://www.kaggle.com/labintsevai/kerasretinanet-wheat-submission
", "rawMarkdown": "To evaluate the different detector solutions, I collect the open solutions on kaggle kernels.It seems that FasterRCNN has the best performance and the efficientdet has the worst performance.\n\n**Two stage detector:**\n\n1.FasterRCNN \nimgage size:1024*1024\nmodel: resnet50\nlb:0.6687\nhttps://www.kaggle.com/pestipeti/pytorch-starter-fasterrcnn-inference\n\n--------------------------------------------------------------------------------------\n\n**One stage detector:**\n\n1- yolov3\nimgage size:608*608\nmodel: Darknet\nlb:0.6504\nhttps://www.kaggle.com/sumitjha19/yolov3-inference-v2-full-code\n\n2- centernet\nimgage size:512*512\nmodel: Resnet18\nlb:0.52\n\nhttps://www.kaggle.com/kyoshioka47/pytorch-centernet-inference-with-tta\n\n3- yolov4\nimgage size:608*608\nmodel: Darknet\nlb:0.6184\n\nhttps://www.kaggle.com/nvnnghia/yolov4-inference\n\n\n\n4- efficientdet\n1) efficientdet0\nimgage size:1024*1024\nmodel: efficientnet b0\nlb:0.5322\n\nhttps://www.kaggle.com/ateplyuk/gwd-starter-efficientdet-inference\n\n2)efficentdet3\nimgage size:512*512\nmodel: efficientnet b3\nlb: \ncv: 0.547\n\n5- retinanet\nimgage size:1024*1024\nmodel: resnet50\nlb: 0.5998\nhttps://www.kaggle.com/labintsevai/kerasretinanet-wheat-submission", "votes": 21 }, { "id": 846682, "postDate": "2020-05-13T23:45:48.047Z", "content": "@alexandersoare I have no enough time to test FasterRCNN. But I think it should get a score over 0.66 . Have a good luck!
", "rawMarkdown": "@alexandersoare I have no enough time to test FasterRCNN. But I think it should get a score over 0.66 . Have a good luck!", "votes": 1 }, { "id": 845719, "postDate": "2020-05-13T11:44:48.947Z", "content": "Thanks for sharing this benchmark @qinhui1999 \nI've tried the FasterRCNN implementation and haven't been able to get above 66 even using the pseudo labelling approach. Wondering if the actual scores have a slightly different approach than as shown in the public kernels.
", "rawMarkdown": "Thanks for sharing this benchmark @qinhui1999 \nI've tried the FasterRCNN implementation and haven't been able to get above 66 even using the pseudo labelling approach. Wondering if the actual scores have a slightly different approach than as shown in the public kernels.", "votes": 1 }, { "id": 853464, "postDate": "2020-05-19T07:56:42.603Z", "content": "FasterRCNN + 2 rounds pseudo labeling + Augmentations = 0.6826
", "rawMarkdown": "FasterRCNN + 2 rounds pseudo labeling + Augmentations = 0.6826", "votes": 2 }, { "id": 846809, "postDate": "2020-05-14T02:31:17.257Z", "content": "as stated in the paper: Global Wheat Head Detection (GWHD) dataset: a large and diverse dataset of high resolution RGB labelled images to develop and benchmark wheat head detection methods, the baseline has a mAP@0.5 of 0.68 using fasterrcnn with resnet34 backbone on 512x512 inputs. Disappointed to hear that efficentdet works badly, which I was planning to try first.
", "rawMarkdown": "as stated in the paper: [Global Wheat Head Detection (GWHD) dataset: a large and diverse dataset of high resolution RGB labelled images to develop and benchmark wheat head detection methods](https://arxiv.org/abs/2005.02162), the baseline has a mAP@0.5 of 0.68 using fasterrcnn with resnet34 backbone on 512x512 inputs. Disappointed to hear that efficentdet works badly, which I was planning to try first. ", "votes": 2 }, { "id": 851847, "postDate": "2020-05-18T00:52:40.400Z", "content": "I think cascadeRcnn will be better .
", "rawMarkdown": "I think cascadeRcnn will be better .", "replies": [ { "id": 851857, "postDate": "2020-05-18T01:07:09.800Z", "content": "@xujingzhao You are right. But there is no baseline of it . So we do not know it's performance.
", "rawMarkdown": "@xujingzhao You are right. But there is no baseline of it . So we do not know it's performance. " }, { "id": 852238, "postDate": "2020-05-18T09:04:20.673Z", "content": "MMdetection is a super good codeBase , but I don't know whether we can install it in notebook.
", "rawMarkdown": "MMdetection is a super good codeBase , but I don't know whether we can install it in notebook." }, { "id": 852255, "postDate": "2020-05-18T09:26:17.087Z", "content": "@xujingzhao Here it is. https://www.kaggle.com/superkevingit/faster-rcnn-with-mmdetection-without-internet
", "rawMarkdown": "@xujingzhao Here it is. https://www.kaggle.com/superkevingit/faster-rcnn-with-mmdetection-without-internet " }, { "id": 852283, "postDate": "2020-05-18T10:02:23.670Z", "content": "Very great , I must try .
", "rawMarkdown": "Very great , I must try ." } ] }, { "id": 853137, "postDate": "2020-05-19T01:50:58.873Z", "rawMarkdown": "", "isDeleted": true }, { "id": 847597, "postDate": "2020-05-14T14:11:21.580Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 847613, "postDate": "2020-05-14T14:18:10.390Z", "rawMarkdown": "", "isDeleted": true } ] } ], "comments": [ { "id": 846682, "author_name": "huiqin", "author_url": "", "post_date": "2020-05-13T23:45:48.047000", "content": "@alexandersoare I have no enough time to test FasterRCNN. But I think it should get a score over 0.66 . Have a good luck!
", "votes": 1, "replies": [] }, { "id": 845719, "author_name": "Alexander Soare", "author_url": "", "post_date": "2020-05-13T11:44:48.947000", "content": "Thanks for sharing this benchmark @qinhui1999 \nI've tried the FasterRCNN implementation and haven't been able to get above 66 even using the pseudo labelling approach. Wondering if the actual scores have a slightly different approach than as shown in the public kernels.
", "votes": 1, "replies": [] }, { "id": 853464, "author_name": "Kaushal Shah", "author_url": "", "post_date": "2020-05-19T07:56:42.603000", "content": "FasterRCNN + 2 rounds pseudo labeling + Augmentations = 0.6826
", "votes": 2, "replies": [] }, { "id": 846809, "author_name": "Tian", "author_url": "", "post_date": "2020-05-14T02:31:17.257000", "content": "as stated in the paper: Global Wheat Head Detection (GWHD) dataset: a large and diverse dataset of high resolution RGB labelled images to develop and benchmark wheat head detection methods, the baseline has a mAP@0.5 of 0.68 using fasterrcnn with resnet34 backbone on 512x512 inputs. Disappointed to hear that efficentdet works badly, which I was planning to try first.
", "votes": 2, "replies": [] }, { "id": 851847, "author_name": "哈尔的移动城堡", "author_url": "", "post_date": "2020-05-18T00:52:40.400000", "content": "I think cascadeRcnn will be better .
", "votes": 0, "replies": [ { "id": 851857, "author_name": "huiqin", "author_url": "", "post_date": "2020-05-18T01:07:09.800000", "content": "@xujingzhao You are right. But there is no baseline of it . So we do not know it's performance.
", "votes": 0, "replies": [] }, { "id": 852238, "author_name": "哈尔的移动城堡", "author_url": "", "post_date": "2020-05-18T09:04:20.673000", "content": "MMdetection is a super good codeBase , but I don't know whether we can install it in notebook.
", "votes": 0, "replies": [] }, { "id": 852255, "author_name": "huiqin", "author_url": "", "post_date": "2020-05-18T09:26:17.087000", "content": "@xujingzhao Here it is. https://www.kaggle.com/superkevingit/faster-rcnn-with-mmdetection-without-internet
", "votes": 0, "replies": [] }, { "id": 852283, "author_name": "哈尔的移动城堡", "author_url": "", "post_date": "2020-05-18T10:02:23.670000", "content": "Very great , I must try .
", "votes": 0, "replies": [] } ] }, { "id": 853137, "author_name": "", "author_url": "", "post_date": "2020-05-19T01:50:58.873000", "content": "", "votes": 0, "replies": [] }, { "id": 847597, "author_name": "", "author_url": "", "post_date": "2020-05-14T14:11:21.580000", "content": "", "votes": 0, "replies": [ { "id": 847613, "author_name": "", "author_url": "", "post_date": "2020-05-14T14:18:10.390000", "content": "", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "845640": "To evaluate the different detector solutions, I collect the open solutions on kaggle kernels.It seems that FasterRCNN has the best performance and the efficientdet has the worst performance.\n\n**Two stage detector:**\n\n1.FasterRCNN \nimgage size:1024*1024\nmodel: resnet50\nlb:0.6687\nhttps://www.kaggle.com/pestipeti/pytorch-starter-fasterrcnn-inference\n\n--------------------------------------------------------------------------------------\n\n**One stage detector:**\n\n1- yolov3\nimgage size:608*608\nmodel: Darknet\nlb:0.6504\nhttps://www.kaggle.com/sumitjha19/yolov3-inference-v2-full-code\n\n2- centernet\nimgage size:512*512\nmodel: Resnet18\nlb:0.52\n\nhttps://www.kaggle.com/kyoshioka47/pytorch-centernet-inference-with-tta\n\n3- yolov4\nimgage size:608*608\nmodel: Darknet\nlb:0.6184\n\nhttps://www.kaggle.com/nvnnghia/yolov4-inference\n\n\n\n4- efficientdet\n1) efficientdet0\nimgage size:1024*1024\nmodel: efficientnet b0\nlb:0.5322\n\nhttps://www.kaggle.com/ateplyuk/gwd-starter-efficientdet-inference\n\n2)efficentdet3\nimgage size:512*512\nmodel: efficientnet b3\nlb: \ncv: 0.547\n\n5- retinanet\nimgage size:1024*1024\nmodel: resnet50\nlb: 0.5998\nhttps://www.kaggle.com/labintsevai/kerasretinanet-wheat-submission", "846682": "@alexandersoare I have no enough time to test FasterRCNN. But I think it should get a score over 0.66 . Have a good luck!", "845719": "Thanks for sharing this benchmark @qinhui1999 \nI've tried the FasterRCNN implementation and haven't been able to get above 66 even using the pseudo labelling approach. Wondering if the actual scores have a slightly different approach than as shown in the public kernels.", "853464": "FasterRCNN + 2 rounds pseudo labeling + Augmentations = 0.6826", "846809": "as stated in the paper: [Global Wheat Head Detection (GWHD) dataset: a large and diverse dataset of high resolution RGB labelled images to develop and benchmark wheat head detection methods](https://arxiv.org/abs/2005.02162), the baseline has a mAP@0.5 of 0.68 using fasterrcnn with resnet34 backbone on 512x512 inputs. Disappointed to hear that efficentdet works badly, which I was planning to try first. ", "851847": "I think cascadeRcnn will be better .", "853137": "", "847597": "" } }