Hello everyone, I have built a YoloV5 model and now I was thinking about trying mmdetection. I have checked their documentation and this notebook https://www.kaggle.com/gauravsingh1/mmdet-eval-fasterrcnn-with-wbf/data
\nBut, I couldn't find answers to my questions. This notebook is using this coco converted dataset
\nhttps://www.kaggle.com/sreevishnudamodaran/vinbigdata-coco-dataset-with-wbf-3x-downscaled
\nMy main question is that shouldn't the bounding boxes be normalized? I copied the notebook, ran it several times, made some changes and the score is always somewhere between 0.010 and 0.030 which is quite low for a FasterRCNN. The default YoloV5 config gives 0.12x…. so I am pretty sure something is wrong, I am just trying to figure out what it is. There aren't that many good mmdetection tutorials out there. (Please don't tell me to check their documentation as I have already done that).
In summary that notebook / my one is using:
\nFasterRCNN with pretrained weights
\ntrained for around 30 epochs on 3x Downscaled images (so around 700-800 x 700-800)
I have the same question.When I visualize the boxes , it seems right.
", "rawMarkdown": "I have the same question.When I visualize the boxes , it seems right.", "votes": 1 }, { "id": 1248519, "postDate": "2021-03-22T16:47:05.427Z", "content": "Hello everyone, I have built a YoloV5 model and now I was thinking about trying mmdetection. I have checked their documentation and this notebook https://www.kaggle.com/gauravsingh1/mmdet-eval-fasterrcnn-with-wbf/data
\nBut, I couldn't find answers to my questions. This notebook is using this coco converted dataset
\nhttps://www.kaggle.com/sreevishnudamodaran/vinbigdata-coco-dataset-with-wbf-3x-downscaled
\nMy main question is that shouldn't the bounding boxes be normalized? I copied the notebook, ran it several times, made some changes and the score is always somewhere between 0.010 and 0.030 which is quite low for a FasterRCNN. The default YoloV5 config gives 0.12x…. so I am pretty sure something is wrong, I am just trying to figure out what it is. There aren't that many good mmdetection tutorials out there. (Please don't tell me to check their documentation as I have already done that).
In summary that notebook / my one is using:
\nFasterRCNN with pretrained weights
\ntrained for around 30 epochs on 3x Downscaled images (so around 700-800 x 700-800)
This code works for square dataset (256x256 for example. In this case dim = 256), you could try to adapt it. The DataFrame is the original one, with original coordinates. Tell me if you have any questions.
\ndef to_coco(df_path, name) :\n\n df = pd.read_csv(df_path) \n annotations = []\n images = []\n obj_count = 0\n for i, image_id in enumerate(df.image_id.unique()) :\n images.append(dict(file_name = f'{image_id}.png', height=dim, width=dim, id=i))\n\n for index, row in df.query(\"image_id == @image_id\").iterrows():\n h_ratio = dim / row['height']\n w_ratio = dim / row['width']\n data_anno = dict(\n image_id=i,\n iscrowd=0,\n id=obj_count,\n category_id=row['class_id'],\n bbox=[\n float(row[\"x_min\"]) * w_ratio,\n float(row[\"y_min\"]) * h_ratio,\n float(row[\"x_max\"]) * w_ratio - float(row[\"x_min\"]) * w_ratio,\n float(row[\"y_max\"]) * h_ratio - float(row[\"y_min\"]) * h_ratio],\n area=(float(row[\"x_max\"]) * w_ratio - float(row[\"x_min\"]) * w_ratio) * (float(row[\"y_max\"]) * h_ratio - float(row[\"y_min\"]) * h_ratio))\n annotations.append(data_anno)\n obj_count += 1\n coco_format_json = dict(\n images=images,\n annotations=annotations,\n categories=cat2label)\n mmcv.dump(coco_format_json, f'{name}')\nAwesome, thanks very much! I will try it and let you know. This isn't writing normalised coordinates to the json file though, so I assume you don't need to normalise them?
", "rawMarkdown": "Awesome, thanks very much! I will try it and let you know. This isn't writing normalised coordinates to the json file though, so I assume you don't need to normalise them?", "votes": 1 }, { "id": 1249833, "postDate": "2021-03-23T14:56:50.240Z", "content": "float(row[\"x_min\"]) * w_ratio You just need to adapt the new coordinates to the new image dimension
I have the same question.When I visualize the boxes , it seems right.
", "votes": 1, "replies": [] }, { "id": 1248670, "author_name": "Matthieu Planté", "author_url": "", "post_date": "2021-03-22T18:34:25.423000", "content": "This code works for square dataset (256x256 for example. In this case dim = 256), you could try to adapt it. The DataFrame is the original one, with original coordinates. Tell me if you have any questions.
\ndef to_coco(df_path, name) :\n\n df = pd.read_csv(df_path) \n annotations = []\n images = []\n obj_count = 0\n for i, image_id in enumerate(df.image_id.unique()) :\n images.append(dict(file_name = f'{image_id}.png', height=dim, width=dim, id=i))\n\n for index, row in df.query(\"image_id == @image_id\").iterrows():\n h_ratio = dim / row['height']\n w_ratio = dim / row['width']\n data_anno = dict(\n image_id=i,\n iscrowd=0,\n id=obj_count,\n category_id=row['class_id'],\n bbox=[\n float(row[\"x_min\"]) * w_ratio,\n float(row[\"y_min\"]) * h_ratio,\n float(row[\"x_max\"]) * w_ratio - float(row[\"x_min\"]) * w_ratio,\n float(row[\"y_max\"]) * h_ratio - float(row[\"y_min\"]) * h_ratio],\n area=(float(row[\"x_max\"]) * w_ratio - float(row[\"x_min\"]) * w_ratio) * (float(row[\"y_max\"]) * h_ratio - float(row[\"y_min\"]) * h_ratio))\n annotations.append(data_anno)\n obj_count += 1\n coco_format_json = dict(\n images=images,\n annotations=annotations,\n categories=cat2label)\n mmcv.dump(coco_format_json, f'{name}')\nAwesome, thanks very much! I will try it and let you know. This isn't writing normalised coordinates to the json file though, so I assume you don't need to normalise them?
", "votes": 1, "replies": [] }, { "id": 1249833, "author_name": "Matthieu Planté", "author_url": "", "post_date": "2021-03-23T14:56:50.240000", "content": "float(row[\"x_min\"]) * w_ratio You just need to adapt the new coordinates to the new image dimension