search

mlzxy / devit

CoRL 2024
https://mlzxy.github.io/devit
MIT License
353 stars 46 forks source link

Few questions regrding costum dataset #44

Open guetaro opened 8 months ago

guetaro commented 8 months ago

Hi,

First of all thanks for this great work.

I would like to run the few shot model on my dataset. I have a single object dataset with tens of tagged images and a lot more of non-tagged images. I would like to perform object detection for this dataset.

I red all guides you've wrote about it but got really confused about how to train & test the model over my dataset.

I do understand I need to organize my data in COCO format, but I do not understand how to do this in this context - what does seen and unseen labels mean? and what does the script 003.ipynb do?

Next, I saw I need to use the script build_prototypes.ipynb to create class prototypes, but this script requires masks for my dataset, which I do not have.

Then, what should I do to train the model?

Thanks in advance,

Roi

Bazenr commented 5 months ago

It is recommended to employ the official 'labelme' to 'coco' conversion script to transform the 'labelme' annotations into the COCO 2014/2017 data format. (https://github.com/labelmeai/labelme/blob/main/examples/instance_segmentation/labelme2coco.py)

For the prototypes, just generate masks from your annotations, name as image_name.mask.suffix, such as 0_20220919_16123200.mask.png

liniceyo commented 3 weeks ago

Hello! Have you solved this problem? I encountered the same issue as you.