Just curious but I would like to know if anyone has used a faster RCNN to handle this task. There is a kernel on using mask RCNN but I believe this task doesn't require instance segmentation.
", "messages": [ { "id": 386323, "postDate": "2018-09-12T15:55:11.640Z", "content": "Just curious but I would like to know if anyone has used a faster RCNN to handle this task. There is a kernel on using mask RCNN but I believe this task doesn't require instance segmentation.
", "rawMarkdown": "Just curious but I would like to know if anyone has used a faster RCNN to handle this task. There is a kernel on using mask RCNN but I believe this task doesn't require instance segmentation.", "votes": 6 }, { "id": 387238, "postDate": "2018-09-14T15:43:30.300Z", "content": "i am using Yolo with pretrained weight on COCO too.DO i really need images with bounding box?i got really bad performance witout them.
", "rawMarkdown": "i am using Yolo with pretrained weight on COCO too.DO i really need images with bounding box?i got really bad performance witout them.", "replies": [ { "id": 387977, "postDate": "2018-09-16T03:09:32.960Z", "content": "I'm not sure. In my case, using images without bounding box improves one percent of the result on test data. I will continue to tune my model.
", "rawMarkdown": "I'm not sure. In my case, using images without bounding box improves one percent of the result on test data. I will continue to tune my model." } ] }, { "id": 386725, "postDate": "2018-09-13T13:00:18.477Z", "content": "@building11_305 do you have plans to discuss your code in the kernels? I'm hunting for a kernel that uses faster rcnn in the training process. Thanks.
", "rawMarkdown": "@building11_305 do you have plans to discuss your code in the kernels? I'm hunting for a kernel that uses faster rcnn in the training process. Thanks.", "replies": [ { "id": 387020, "postDate": "2018-09-14T06:27:30.023Z", "content": "\n\n\nshiv wrote
\n \n\n\n@building11_305 do you have plans to discuss your code in the kernels? I'm hunting for a kernel that uses faster rcnn in the training process. Thanks.
\n
I'm testing the idea of https://github.com/rbgirshick/py-faster-rcnn/issues/231. BUT I get poor results of rpn right now. If I succeed, I will share my idea. You can refer to tensorpack https://github.com/tensorpack/tensorpack/tree/master/examples/FasterRCNN. VERY POWERFUL implementation.
", "rawMarkdown": "\n> **shiv wrote**\n> \n> > @building11_305 do you have plans to discuss your code in the kernels? I'm hunting for a kernel that uses faster rcnn in the training process. Thanks.\n\nI'm testing the idea of https://github.com/rbgirshick/py-faster-rcnn/issues/231. BUT I get poor results of rpn right now. If I succeed, I will share my idea. You can refer to tensorpack https://github.com/tensorpack/tensorpack/tree/master/examples/FasterRCNN. VERY POWERFUL implementation. ", "votes": 2 } ] }, { "id": 386600, "postDate": "2018-09-13T08:05:18.810Z", "content": "I'm using faster rcnn. BUT I still don't have a good idea to use the image without bounding box. Right Now I just use the image with bounding boxes to train model.
", "rawMarkdown": "I'm using faster rcnn. BUT I still don't have a good idea to use the image without bounding box. Right Now I just use the image with bounding boxes to train model.", "replies": [ { "id": 386707, "postDate": "2018-09-13T12:22:31.623Z", "content": "I am using Faster R-CNN ResNet101 pretrained on COCO.
\nI tried Prune's answer in \"https://stackoverflow.com/questions/46589962/should-i-include-negative-examples-for-tensorflow-object-detection-api?rq=1\" to use images without bounding boxes, but got a very poor accuracy.
Did you do any sampling for the images? I'm trying a RetinaNet, but I have very poor performance unless I sample more abnormal images than normal images.
", "rawMarkdown": "Did you do any sampling for the images? I'm trying a RetinaNet, but I have very poor performance unless I sample more abnormal images than normal images." }, { "id": 386765, "postDate": "2018-09-13T14:52:45.493Z", "content": "I randomly sampled no opacity images. The ratio of sampled normal and abnormal will be same as that of original entire dataset( normal:abnormal = 8525:11500 ).
", "rawMarkdown": "I randomly sampled no opacity images. The ratio of sampled normal and abnormal will be same as that of original entire dataset( normal:abnormal = 8525:11500 )." }, { "id": 387027, "postDate": "2018-09-14T06:40:50.357Z", "content": "Maybe this is a good start https://github.com/rbgirshick/py-faster-rcnn/issues/231. But it seems that we need to rebalance the ratio of fg and bg.
", "rawMarkdown": "Maybe this is a good start https://github.com/rbgirshick/py-faster-rcnn/issues/231. But it seems that we need to rebalance the ratio of fg and bg.", "votes": 1 } ] }, { "id": 386419, "postDate": "2018-09-12T20:19:04.563Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 387238, "author_name": "天马流星拳", "author_url": "", "post_date": "2018-09-14T15:43:30.300000", "content": "i am using Yolo with pretrained weight on COCO too.DO i really need images with bounding box?i got really bad performance witout them.
", "votes": 0, "replies": [ { "id": 387977, "author_name": "building11_301", "author_url": "", "post_date": "2018-09-16T03:09:32.960000", "content": "I'm not sure. In my case, using images without bounding box improves one percent of the result on test data. I will continue to tune my model.
", "votes": 0, "replies": [] } ] }, { "id": 386725, "author_name": "shiv", "author_url": "", "post_date": "2018-09-13T13:00:18.477000", "content": "@building11_305 do you have plans to discuss your code in the kernels? I'm hunting for a kernel that uses faster rcnn in the training process. Thanks.
", "votes": 0, "replies": [ { "id": 387020, "author_name": "building11_301", "author_url": "", "post_date": "2018-09-14T06:27:30.023000", "content": "\n\n\nshiv wrote
\n \n\n\n@building11_305 do you have plans to discuss your code in the kernels? I'm hunting for a kernel that uses faster rcnn in the training process. Thanks.
\n
I'm testing the idea of https://github.com/rbgirshick/py-faster-rcnn/issues/231. BUT I get poor results of rpn right now. If I succeed, I will share my idea. You can refer to tensorpack https://github.com/tensorpack/tensorpack/tree/master/examples/FasterRCNN. VERY POWERFUL implementation.
", "votes": 2, "replies": [] } ] }, { "id": 386600, "author_name": "building11_301", "author_url": "", "post_date": "2018-09-13T08:05:18.810000", "content": "I'm using faster rcnn. BUT I still don't have a good idea to use the image without bounding box. Right Now I just use the image with bounding boxes to train model.
", "votes": 0, "replies": [ { "id": 386707, "author_name": "T Nakamura", "author_url": "", "post_date": "2018-09-13T12:22:31.623000", "content": "I am using Faster R-CNN ResNet101 pretrained on COCO.
\nI tried Prune's answer in \"https://stackoverflow.com/questions/46589962/should-i-include-negative-examples-for-tensorflow-object-detection-api?rq=1\" to use images without bounding boxes, but got a very poor accuracy.
Did you do any sampling for the images? I'm trying a RetinaNet, but I have very poor performance unless I sample more abnormal images than normal images.
", "votes": 0, "replies": [] }, { "id": 386765, "author_name": "T Nakamura", "author_url": "", "post_date": "2018-09-13T14:52:45.493000", "content": "I randomly sampled no opacity images. The ratio of sampled normal and abnormal will be same as that of original entire dataset( normal:abnormal = 8525:11500 ).
", "votes": 0, "replies": [] }, { "id": 387027, "author_name": "building11_301", "author_url": "", "post_date": "2018-09-14T06:40:50.357000", "content": "Maybe this is a good start https://github.com/rbgirshick/py-faster-rcnn/issues/231. But it seems that we need to rebalance the ratio of fg and bg.
", "votes": 1, "replies": [] } ] }, { "id": 386419, "author_name": "", "author_url": "", "post_date": "2018-09-12T20:19:04.563000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "386323": "Just curious but I would like to know if anyone has used a faster RCNN to handle this task. There is a kernel on using mask RCNN but I believe this task doesn't require instance segmentation.", "387238": "i am using Yolo with pretrained weight on COCO too.DO i really need images with bounding box?i got really bad performance witout them.", "386725": "@building11_305 do you have plans to discuss your code in the kernels? I'm hunting for a kernel that uses faster rcnn in the training process. Thanks.", "386600": "I'm using faster rcnn. BUT I still don't have a good idea to use the image without bounding box. Right Now I just use the image with bounding boxes to train model.", "386419": "" } }