ACE0 + UGCR v5 + UAR Head — 改进版(2026-05-27 snapshot)

基于 ACE0 (ECCV 2024 Oral) 的两项扩展,提升场景坐标回归的可靠性与光照鲁棒性。

下载与解压

sudo apt install zstd  # 或 brew install zstd

# 下载 (1.8 GB)
huggingface-cli download taopeng/ace0-ugcr-uar ace0-ugcr-uar.tar.zst --local-dir .

# 解压
tar --use-compress-program=unzstd -xf ace0-ugcr-uar.tar.zst

解压得到 20260415_s3_upload/ 目录 (~6.4 GB)。

主要方法

  • UGCR v5 — 不确定性引导坐标回归 (log1p(error) Gaussian NLL + 分阶段训练 85% warmup)
  • UAR Head — 不确定性感知精修 Head (σ 驱动空间注意力 + 坐标精修残差)
  • Combined v4 — UGCR + UAR 联合

主要结果

Metric Baseline UGCR v5 UAR Head
7-Scenes Acc @5cm/5° avg 88.9% 89.6% 90.3%
7-Scenes Acc @2cm/2° heads 89.3% 96.1% 100.0%
Indoor6 high-conf pose 数 (mean) 146.5 224.3 (+53%) 209.7 (+43%)

详见 experiments/results/final_results.mdpaper_ready_report.mdFINAL_RESULTS_EXTENDED.md

目录

  • idea2_ugcr/ — UGCR + UAR Head 源码(含 log1p + σ.detach + UAR refine_net)
  • idea1_igda/ — IGDA 系列源码
  • baseline_ace0/ — ACE0 baseline (改动版本)
  • experiments/scripts/ — 训练/评估/出图脚本(含 run_ugcr_v5_full.sh, eval_multi_threshold.py, make_figures.py
  • experiments/results/ — 全部方法的 pose 结果 + 评估输出 + figures_v2 (2cm/2° boxplot + recall curves)

Companion repo

完整 baseline 数据(含 7-Scenes/Aachen/Indoor6 raw 数据集): taopeng/ace0-rgbt-vl

引用

  • Brachmann et al., Scene Coordinate Reconstruction (ACE0), ECCV 2024 Oral. arXiv:2404.14351
  • Kendall & Gal, What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?, NeurIPS 2017. arXiv:1703.04977
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Papers for taopeng/ace0-ugcr-uar