I wonder if anyone tried to use SSD or faster rcnn with this dataset?\nI tried to use 8 layer deep ssd , but it seems that it's not enough. \nDoes anyone have any suggestion for start poing? \nThank you!
", "messages": [ { "id": "178663", "postDate": "04/28/2017 15:16:24", "content": "I wonder if anyone tried to use SSD or faster rcnn with this dataset?\nI tried to use 8 layer deep ssd , but it seems that it's not enough. \nDoes anyone have any suggestion for start poing? \nThank you!
", "rawMarkdown": "I wonder if anyone tried to use SSD or faster rcnn with this dataset?\nI tried to use 8 layer deep ssd , but it seems that it's not enough. \nDoes anyone have any suggestion for start poing? \nThank you!", "votes": null }, { "id": "178742", "postDate": "04/28/2017 19:21:44", "content": "What accuracy are you achieving?
", "rawMarkdown": "What accuracy are you achieving?", "votes": null }, { "id": "179254", "postDate": "04/30/2017 20:04:07", "content": "Hey Matvey, I am gonna use faster-RCNN but need ground-truth bounding boxes. Do you have boxes already? Maybe we could work together? I've quite a bit of experience on object detection from other work.
", "rawMarkdown": "Hey Matvey, I am gonna use faster-RCNN but need ground-truth bounding boxes. Do you have boxes already? Maybe we could work together? I've quite a bit of experience on object detection from other work.", "votes": null }, { "id": "182589", "postDate": "05/14/2017 18:51:53", "content": "Hi all again. I used ssd to detect lions, and then I used k-means to sort them, using bounding box diogonal. Unfortunatelly the error is still very high. I have bboxes for 300 images, but i crop them for 500x500 pixels. If someone want to join me in cuting boxes - send add e-mail here
", "rawMarkdown": "Hi all again. I used ssd to detect lions, and then I used k-means to sort them, using bounding box diogonal. Unfortunatelly the error is still very high. I have bboxes for 300 images, but i crop them for 500x500 pixels. If someone want to join me in cuting boxes - send add e-mail here", "votes": null }, { "id": "184745", "postDate": "05/22/2017 22:49:06", "content": "For SSD, each feature map layer is responsible for a specific scale of the image you are trying to detect. As you go deeper, the feature maps get smaller, which allow the network to detect smaller images. I'd imagine for this task, you would want to use less layers, since the sea lions are more or less the same scale.
", "rawMarkdown": "For SSD, each feature map layer is responsible for a specific scale of the image you are trying to detect. As you go deeper, the feature maps get smaller, which allow the network to detect smaller images. I'd imagine for this task, you would want to use less layers, since the sea lions are more or less the same scale.", "votes": null }, { "id": "184884", "postDate": "05/23/2017 12:14:38", "content": "The are actually not. The variation of scale is quiet big.
", "rawMarkdown": "The are actually not. The variation of scale is quiet big.", "votes": null }, { "id": "185249", "postDate": "05/24/2017 16:57:52", "content": "Any suggestions on how to adjust a cnn model to deal with the variation of scale? To me, a lot times that I can distinguish which class the sea lion is are based on their size, I don't understand how the cnn can learn things accurately given the variation of scales.
", "rawMarkdown": "Any suggestions on how to adjust a cnn model to deal with the variation of scale? To me, a lot times that I can distinguish which class the sea lion is are based on their size, I don't understand how the cnn can learn things accurately given the variation of scales.", "votes": null }, { "id": "186254", "postDate": "05/27/2017 08:57:59", "content": "Yes, this scaling thing is a key question for me - see my heat maps post - I would like to collaborate with anyone who can help me crack this scaling problem.
", "rawMarkdown": "Yes, this scaling thing is a key question for me - see my heat maps post - I would like to collaborate with anyone who can help me crack this scaling problem.", "votes": null }, { "id": "193448", "postDate": "06/16/2017 13:31:15", "content": "I have tried SSD, it didn't work well on crowded images.
", "rawMarkdown": "I have tried SSD, it didn't work well on crowded images.", "votes": null } ], "comments": [ { "id": 178742, "author_name": "timjoseph", "author_url": "", "post_date": "04/28/2017 19:21:44", "content": "What accuracy are you achieving?
", "votes": null, "replies": [] }, { "id": 179254, "author_name": "chandrachud", "author_url": "", "post_date": "04/30/2017 20:04:07", "content": "Hey Matvey, I am gonna use faster-RCNN but need ground-truth bounding boxes. Do you have boxes already? Maybe we could work together? I've quite a bit of experience on object detection from other work.
", "votes": null, "replies": [] }, { "id": 182589, "author_name": "hagorms", "author_url": "", "post_date": "05/14/2017 18:51:53", "content": "Hi all again. I used ssd to detect lions, and then I used k-means to sort them, using bounding box diogonal. Unfortunatelly the error is still very high. I have bboxes for 300 images, but i crop them for 500x500 pixels. If someone want to join me in cuting boxes - send add e-mail here
", "votes": null, "replies": [] }, { "id": 184745, "author_name": "walterwangnyc", "author_url": "", "post_date": "05/22/2017 22:49:06", "content": "For SSD, each feature map layer is responsible for a specific scale of the image you are trying to detect. As you go deeper, the feature maps get smaller, which allow the network to detect smaller images. I'd imagine for this task, you would want to use less layers, since the sea lions are more or less the same scale.
", "votes": null, "replies": [ { "id": 184884, "author_name": "asanakoev", "author_url": "", "post_date": "05/23/2017 12:14:38", "content": "The are actually not. The variation of scale is quiet big.
", "votes": null, "replies": [] }, { "id": 185249, "author_name": "", "author_url": "", "post_date": "05/24/2017 16:57:52", "content": "Any suggestions on how to adjust a cnn model to deal with the variation of scale? To me, a lot times that I can distinguish which class the sea lion is are based on their size, I don't understand how the cnn can learn things accurately given the variation of scales.
", "votes": null, "replies": [] }, { "id": 186254, "author_name": "jinkos", "author_url": "", "post_date": "05/27/2017 08:57:59", "content": "Yes, this scaling thing is a key question for me - see my heat maps post - I would like to collaborate with anyone who can help me crack this scaling problem.
", "votes": null, "replies": [] } ] }, { "id": 193448, "author_name": "larion", "author_url": "", "post_date": "06/16/2017 13:31:15", "content": "I have tried SSD, it didn't work well on crowded images.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "178663": "I wonder if anyone tried to use SSD or faster rcnn with this dataset?\nI tried to use 8 layer deep ssd , but it seems that it's not enough. \nDoes anyone have any suggestion for start poing? \nThank you!", "178742": "What accuracy are you achieving?", "179254": "Hey Matvey, I am gonna use faster-RCNN but need ground-truth bounding boxes. Do you have boxes already? Maybe we could work together? I've quite a bit of experience on object detection from other work.", "182589": "Hi all again. I used ssd to detect lions, and then I used k-means to sort them, using bounding box diogonal. Unfortunatelly the error is still very high. I have bboxes for 300 images, but i crop them for 500x500 pixels. If someone want to join me in cuting boxes - send add e-mail here", "184745": "For SSD, each feature map layer is responsible for a specific scale of the image you are trying to detect. As you go deeper, the feature maps get smaller, which allow the network to detect smaller images. I'd imagine for this task, you would want to use less layers, since the sea lions are more or less the same scale.", "184884": "The are actually not. The variation of scale is quiet big.", "185249": "Any suggestions on how to adjust a cnn model to deal with the variation of scale? To me, a lot times that I can distinguish which class the sea lion is are based on their size, I don't understand how the cnn can learn things accurately given the variation of scales.", "186254": "Yes, this scaling thing is a key question for me - see my heat maps post - I would like to collaborate with anyone who can help me crack this scaling problem.", "193448": "I have tried SSD, it didn't work well on crowded images." }, "source": "meta" }