DAMO-YOLO: A High‑Efficiency, High‑Accuracy Object Detection Framework

Object detection aims to locate and classify objects within images, videos, or point clouds, providing bounding boxes and class labels. It underpins many computer‑vision applications such as autonomous driving, port management, intrusion detection, and face recognition, making it a highly competitive research area.

Existing detection frameworks often suffer from three main pain points: limited model scale flexibility, weak multi‑scale detection especially for small objects, and sub‑optimal speed‑accuracy trade‑offs.

DAMO‑YOLO addresses these issues with three distinct technical advantages. First, it integrates a self‑developed NAS method (MAE‑NAS) that enables low‑cost, latency‑or FLOPS‑aware model customization without requiring real data or full training. Second, it adopts Efficient RepGFPN together with a HeavyNeck design to substantially improve multi‑scale feature fusion, allocating nearly half of the model’s FLOPS to the neck for balanced performance. Third, it introduces a universal distillation technique that works across tiny, medium, and large models, improving accuracy without adding inference overhead and remaining robust to heterogeneous teacher‑student architectures.

The MAE‑NAS approach treats a network as a continuous‑state information system, maximizing its entropy by modeling vertices (features) and edges (operators) as a graph. Entropy is approximated via feature‑map variance, guiding the search toward architectures that maximize expressive power under given latency or FLOPS budgets. The search is zero‑shot, requiring only a few minutes on a CPU.

Efficient RepGFPN refines the original GFPN by assigning different channel counts to each scale, removing costly up‑sampling links in the queen‑fusion module, and fixing the number of fusion nodes. A lightweight fusion block with re‑parameterization and multi‑layer aggregation further enhances feature merging while preserving parallel efficiency. The head is simplified to a single linear projection (ZeroHead), shifting most computation to the neck (HeavyNeck).

For model distillation, DAMO‑YOLO aligns student features to the teacher, normalizes them with bias‑free batch‑norm, and applies a dynamically decaying loss weight to avoid hindering the student’s classification branch. This scheme improves all model scales and remains effective for heterogeneous teacher‑student pairs.

Empirically, DAMO‑YOLO achieves 20‑40% speedup, 15‑50% reduction in FLOPS, and 6‑50% fewer parameters compared to state‑of‑the‑art detectors at comparable accuracy, with notable gains on both small and large objects. The models are available on ModelScope and GitHub, ready for inference and training with minimal configuration.

Future plans include expanding deployment tools, adding more application examples (e.g., drone‑based small‑object detection, rotated‑box detection), and releasing nano‑scale models for edge devices as well as large models for cloud scenarios.