When even nano-YOLO is too bloated, try 250K weights

What it does FastestDet is a single-scale, anchor-free object detector that tries to shrink the model footprint until it nearly vanishes. It runs a 352×352 image through roughly 250K parameters and targets low-end ARM boards like the RK3568 and Snapdragon 835 via NCNN or RKNN, without requiring a GPU. Training uses standard Darknet-style Yolo datasets, and the model exports to ONNX and TorchScript for deployment.

The interesting bit Instead of a multi-head feature pyramid, the network relies on one detector head and a “cross grid multiple candidate targets” strategy to recover recall from its tiny backbone. The author also lists dynamic positive/negative sample allocation as an improvement, though the README treats it as a bullet point rather than an explanation.

Key highlights

• Anchor-free with a single feature-map head; the author says this simplifies post-processing compared with earlier Yolo-fastest variants.
• Benchmarked at 23.5 ms (quad-core) and 70.6 ms (single-core) on a 2.0 GHz ARM Cortex-A55 via NCNN, with 25.3% mAP@0.5 on COCO.
• Supports NCNN, RKNN, ONNX Runtime, and TorchScript across Linux, Android, and x86.
• Loss function updated in mid-2022 to IOU-aware smooth L1, which the author credits with a 0.7 mAP boost.

Caveats

• Accuracy is firmly in the “proof of life” category: 13.0% mAP@0.5:0.95 and 25.3% mAP@0.5 on COCO, well behind larger nano detectors like YOLOX-nano or NanoDet.
• The single-core ARM runtime (70.62 ms) is actually slightly slower than the predecessor Yolo-fastestv2 (68.9 ms), so the claimed ~10% speedup appears to come from multi-core or older Yolo-fastestv1.1 baselines.
• The README’s last noted update is July 2022, and several sections read like raw notes rather than maintained documentation.

Verdict Grab it if you are prototyping on a low-end ARM board and need any object detection at all, not good object detection. Skip it if you need reliable COCO-grade accuracy or a project with active maintenance.