// the find
tinyvision/DAMO-YOLO
DAMO-YOLO: a fast and accurate object detection method with some new techs, including NAS backbones, efficient RepGFPN, ZeroHead, AlignedOTA, and distillation enhancement.
DAMO-YOLO is Alibaba's object detection model from 2022, competing in the YOLO family with NAS-searched backbones, a reparameterized FPN neck (RepGFPN), and a lightweight detection head. It targets the accuracy/latency tradeoff on T4 GPUs and Intel CPUs, with ONNX and TensorRT export paths built in. Aimed at CV practitioners who want a strong pretrained baseline without building their own detection pipeline from scratch.
The NAS backbone search is genuinely useful — you can run MAE-NAS under your own latency budget and drop in a custom backbone via a text file, which is more practical than hand-designing one. The distillation chain (S teaches T, M teaches S, etc.) is well-executed and the mAP gains are real: DAMO-YOLO-T* hits 43.6 mAP at 2.78ms T4 latency, which is competitive. The 701-category pretrained model trained on COCO + Objects365 + OpenImage gives a meaningful head start for downstream fine-tuning, as shown by the VisDrone results. TensorRT Int8 partial quantization with a documented 19% speedup at 0.3% accuracy loss is a concrete, usable tradeoff rather than a vague claim.
The repo has been effectively dead since May 2024, with the last real development burst in early 2023 — YOLO has moved on substantially (YOLOv9, YOLO11) and the benchmarks here are no longer competitive at the top end. The installation instructions pin PyTorch 1.7 and CUDA 10.2, which are years out of date and will cause environment headaches on any modern machine. Google Drive download links for the legacy models are broken (marked as --), and model weights live on Alibaba OSS which may have availability issues outside China. There is no Python package installable via pip — you clone the repo, manually set PYTHONPATH, and pray your conda environment resolves; this is friction that rules it out for anything production-adjacent.