![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F357928%2Fa642bb1be1706a74dd71fcd63331877e%2FScreen%20Shot%202020-05-15%20at%2011.22.42%20PM.png?generation=1589610207287039&alt=media\")
I am new to object detection problems and trying to train a YoloV3 model for this problem.\nI did a describe on dimensions boxes in training data and it seems to have a lot of variation. \n
Additionally I loaded some images with Anchor boxes and seems like there are a lot of different dimensions.\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F357928%2Fabe249096e439cae1caa6d703c777d72%2FScreen%20Shot%202020-05-15%20at%2011.24.20%20PM.png?generation=1589610280210574&alt=media\")
I know that the CNN will learn and assign different sizes to the output anchor boxes on prediction but I am confused on how to select the total number of anchor boxes. Looking at the above image, I want to select a big number like > 100 so that my IoU is not very far off and specific box sizes can be found but at the same I fear that it may make my model really hard to train.
\n\nWhat is a good anchor box count generally and how is it determined?
", "messages": [ { "id": "849879", "postDate": "05/16/2020 06:26:39", "content": "I am new to object detection problems and trying to train a YoloV3 model for this problem.\nI did a describe on dimensions boxes in training data and it seems to have a lot of variation. \n
Additionally I loaded some images with Anchor boxes and seems like there are a lot of different dimensions.\n
I know that the CNN will learn and assign different sizes to the output anchor boxes on prediction but I am confused on how to select the total number of anchor boxes. Looking at the above image, I want to select a big number like > 100 so that my IoU is not very far off and specific box sizes can be found but at the same I fear that it may make my model really hard to train.
\n\nWhat is a good anchor box count generally and how is it determined?
", "rawMarkdown": "I am new to object detection problems and trying to train a YoloV3 model for this problem.\nI did a describe on dimensions boxes in training data and it seems to have a lot of variation. \n\n \nAdditionally I loaded some images with Anchor boxes and seems like there are a lot of different dimensions.\n\n \nI know that the CNN will learn and assign different sizes to the output anchor boxes on prediction but I am confused on how to select the total number of anchor boxes. Looking at the above image, I want to select a big number like > 100 so that my IoU is not very far off and specific box sizes can be found but at the same I fear that it may make my model really hard to train. \n\nWhat is a good anchor box count generally and how is it determined?", "votes": null }, { "id": "850325", "postDate": "05/16/2020 14:19:21", "content": "you have to defined threshold point of IOU\nlike : boxes = boxes[scores >= detection_threshold]
", "rawMarkdown": "you have to defined threshold point of IOU\nlike : boxes = boxes[scores >= detection_threshold]", "votes": null } ], "comments": [ { "id": 850325, "author_name": "ajax0564", "author_url": "", "post_date": "05/16/2020 14:19:21", "content": "you have to defined threshold point of IOU\nlike : boxes = boxes[scores >= detection_threshold]
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "849879": "I am new to object detection problems and trying to train a YoloV3 model for this problem.\nI did a describe on dimensions boxes in training data and it seems to have a lot of variation. \n\n \nAdditionally I loaded some images with Anchor boxes and seems like there are a lot of different dimensions.\n\n \nI know that the CNN will learn and assign different sizes to the output anchor boxes on prediction but I am confused on how to select the total number of anchor boxes. Looking at the above image, I want to select a big number like > 100 so that my IoU is not very far off and specific box sizes can be found but at the same I fear that it may make my model really hard to train. \n\nWhat is a good anchor box count generally and how is it determined?", "850325": "you have to defined threshold point of IOU\nlike : boxes = boxes[scores >= detection_threshold]" }, "source": "meta" }