custom dataset folder structure

[BalasubramanyamEvani]

Hello,

Thank you for open-sourcing your implementation

As I understand. In order to finetune super point on a custom dataset, I'd need to follow the below steps:

1. Download a pre-trained detector

2. Use the above detector to generate detections (pseudo ground truth) on the custom dataset (Step 2 in README) - self-supervised step?

3. If the pseudo-ground truth makes sense then

4. Train SuperPoint on a custom dataset (Step 6 in README)

If the above understanding is correct then I was wondering what the dataset folder structure should look like. As I understand MS COCO would have folders for Train and Val. Is the export script scanning for all images in all the folders?

Also, If you prefer communication via mail I'd be happy to do so. Since this is a question and not an actual issue with the codebase.

[rpautrat]

Hi,

Yes your pipeline would be correct. I would also add the following between steps 1 and 2:

1.5 Create a dataloader file in superpoint/datasets by copying the superpoint/datasets/coco.py file. You can then update the links, default parameters and change the structure of the image folder as you wish.

In the case of COCO, there was a single folder with all images, and the dataloader was automatically splitting it into train/val parts. The export script was then scanning all the images of this folder.