MicroSight

MicroSight is an efficient transformer-based middleware for improving small object detection via multi-perspective feature extraction and collaboration. This v1.0.0 release provides the first public, reproducible code drop, including both a DETR-style implementation and an MMDetection/Co-DETR-style implementation. Highlights Two reference implementations included: Mi-DETR : DETR-style training/evaluation entrypoints Mi-CODETR : MMDetection/Co-DETR-style project with configs/tools/projects layout Improved reproducibility and documentation: Expanded top-level README with environment guidance, installation, data preparation, training and evaluation commands Added a top-level requirements.txt to reference subproject dependencies Research artifact citation support: Added CITATION.cff Added a "Manuscript, Preprint, and Citation" section in README Licensing: Added top-level LICENSE (MIT) Added Mi-DETR/LICENSE (Apache-2.0) to match DETR licensing expectations Getting Started See README.md for setup instructions. Choose one of the two workflows: Mi-DETR/ for a lightweight DETR-style pipeline Mi-CODETR/ for MMDetection-style training and configs Notes / Known Caveats Mi-DETR and Mi-CODETR have different dependency stacks; install per subproject for best compatibility. The preprint URL/DOI fields are currently placeholders ( TBD ) and should be updated once the preprint link is finalized.