search

destiny301 / dpr

[WACV2024] Patch-based Selection and Refinement for Early Object Detection
17 stars 3 forks source link

How to set up dataset for training? #1

Open QuangDucc opened 7 months ago

QuangDucc commented 7 months ago

Hello, Thank you for sharing the code.

I have a question about the dataset Bdd100k you were used in your experiment. As you describe, the training and valid dataset will following this format:

root/ ├──images/ │ ├── train/ │ │ ├── 000001.jpg │ │ ├── 000002.jpg │ │ ├── ......

├──masks/ │ ├── val/ │ │ ├── 000001.png │ │ ├── 000002.png │ │ ├── ......

I downloaded the dataset but no part of it following this format. Can you explain more details the part of Bdd100k was used and how to setup the dataset for training.

Thank you in advance for your support.

destiny301 commented 7 months ago

Hi,

Generally speaking, if you want to generate the masks with object detection labels (bounding boxes), you need to stack all boxes together for each image (object=1, background=0). If you want to generate the masks with segmentation labels, you can simply set all nonzero values to 1.

Btw, for the segmentation label, I mean the colormap labels for instance segmentation.

If you have more questions, please let me know.

Chenyali07 commented 7 months ago

I used the json_to_dataset.py file that comes with labelme to generate a json folder of .jpg files, which contains three .png files. Which .png file should be used for training? Or should I use some other way to generate .png images? I'm really looking forward to your help answering my questions. (The following three pictures are three .png images generated by labelme) image image image

destiny301 commented 7 months ago

I used the json_to_dataset.py file that comes with labelme to generate a json folder of .jpg files, which contains three .png files. Which .png file should be used for training? Or should I use some other way to generate .png images? I'm really looking forward to your help answering my questions. (The following three pictures are three .png images generated by labelme) image image image

The second one looks reasonable. But the label should not be RGB image, you need to transfer it into grayscale. (i.e. extract the red channel of the image)

Chenyali07 commented 7 months ago

OK! I will do some processing on the dataset again.