reth0608/CNN-Transformer_Hybrid ? reverse-engineered prompt

Build me a lightweight object detection project for KITTI that compares a simple CNN detector against a hybrid CNN plus Transformer version.

I want it to load KITTI images and labels, train on a YOLO style grid setup, and support a baseline model plus a hybrid model that only sends uncertain detections through a refinement step. It should include training, evaluation, and inference scripts, plus a way to compare the two models using mAP at 0.5 and save example visualizations on images.

Please make it easy to run on a small subset first, then scale up to the full KITTI split. I’d like config files for the different training setups, including a quick MVP version and a better tuned version. If anything is unclear, look up the current docs online if you need to, but keep the project practical and lightweight rather than research heavy.