LibreTinyFormerx-visdrone

TinyFormer-X-PBM detection weights, repackaged for LibreYOLO. TinyFormer ("TinyFormer: Preserving Tiny Objects in YOLO-DETR Hybrid Real-time Detectors", arXiv:2605.25046) is a DEIMv2-derived YOLO-DETR hybrid specialised for tiny objects via a Spatial Semantic Adapter and a 4-scale Parallel Bi-fusion neck. This size runs on a finetuned DINOv3 ViT-S/16+ tower.

Reported VisDrone2019 val AP (upstream protocol): 34.7.

This is the official VisDrone2019 finetune (10 classes: pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus, motor), initialised from the COCO checkpoint. The reported AP follows the upstream repository's VisDrone conversion and evaluation protocol and is not comparable to COCO numbers.

Usage

from libreyolo import LibreYOLO

model = LibreYOLO("LibreTinyFormerx-visdrone.pt")   # auto-downloads from this repo
results = model.predict("image.jpg")

Source

Derived from mmpmmpmmpjosh/TinyFormer at commit 9075d9f, sourced from the official Google Drive checkpoint mirror. Copyright (C) 2026 AICVlab, National Yang Ming Chiao Tung University (NYCU). The TinyFormer codebase is licensed under the Apache License, Version 2.0 (DEIM/DEIMv2 lineage).

The backbone parameters derive from facebookresearch/dinov3 (a finetuned DINOv3 ViT-S/16+ tower) and are licensed under the DINOv3 License. Because this checkpoint bundles DINOv3-derived parameters, the combined distribution is governed by the terms of both licenses.

Modifications

Checkpoint metadata wrapping only — learned parameters are unchanged (the EMA weights of the released solver checkpoint). The wrapper adds model_family, size, nc, and names fields so the unified LibreYOLO() factory routes correctly without filename heuristics. See weights/convert_tinyformer_weights.py in the LibreYOLO source repository.

License

Dual-licensed: Apache License 2.0 (TinyFormer/DEIMv2 components) and the DINOv3 License (backbone parameters). The combined LICENSE file in this repository contains both license texts. See also NOTICE for attribution detail.