Tips for Using PE Spatial on TACO Dataset (Object Detection)

Hi, and thank you for releasing this great work.

I'm working on adapting **PE Spatial** for object detection on the [TACO dataset](https://github.com/pedropro/TACO) (Trash Annotations in Context), which has 60+ object classes of litter in real-world scenes.

I’ve reviewed the paper and code, but I’d appreciate some pointers on how best to proceed with the following:

1. Is there a preferred pipeline or example config for finetuning PEspatial on a custom dataset like TACO?
2. Should I extract intermediate features from a specific layer, or is the aligned output of PEspatial already suitable for detection with standard decoders (e.g., ViTDet + Mask R-CNN)? Is it possible to use YOLO?
3. Any recommendations on input resolution, training schedule, or data augmentations that transfer well from COCO to a smaller dataset like TACO?
4. Would it be better to freeze the backbone or allow full finetuning given the domain shift?

If there's a public example or finetuning script available (even for COCO), that would help me a lot in getting started.

Thanks again for the release, looking forward to exploring PE further!