kpshinnik/repro-infogeo-bundle

Reproduction bundle — InfoGeo (ICML 2026, arXiv 2605.07099)

Independent reproduction of InfoGeo: Information-Theoretic Object-Centric Learning for Cross-View Generalizable UAV Geo-Localization (OpenReview v6xrfEjNQv).

What this bundle contains

scripts/
  mechanisms.py     Claim 1 — faithful re-implementation + property tests of
                    CACS (Eq.7-9), CSRR (Eq.10-13), OCVA (Eq.14-17), IB (Eq.18-20)
  tables_check.py   Claims 2-5 — arithmetic-consistency checks of Tables 1,2,3,5
  toy_retrieval.py  R@1/AP/SDM/Dis@1 metric harness + toy Lstruct ablation (directional)
  make_figs.py      generates the poster figures from the logged JSON
outputs/
  mechanisms_result.json     7/7 mechanism tests pass
  tables_check_result.json   14/14 table-arithmetic checks pass
  toy_retrieval_result.json  metric harness + toy ablation (+2.78 R@1 from Lstruct)
poster/
  poster.html          print poster (60x36in, posterly landscape_4col; all gates pass)
  poster_embed.html    logbook embed (pinned on the Executive summary page)
  poster_preview.png / .pdf
  images/              figures rendered from the logged numbers
infogeo.txt            extracted paper text

How to rerun

python3 -m pip install torch numpy scipy matplotlib pillow
cd scripts
python3 mechanisms.py      # -> outputs/mechanisms_result.json   (asserts 7/7)
python3 tables_check.py    # -> outputs/tables_check_result.json (asserts 14/14)
python3 toy_retrieval.py   # -> outputs/toy_retrieval_result.json (~1 min, CPU)
python3 make_figs.py       # regenerate poster figures

Scope (honest)

• Claim 1 (method/mechanism): reproduced. The three modules and the IB objective are re-implemented from the equations and pass 7/7 property tests, cross-checked against the authors' released code (github.com/HRT00/Official_InfoGeo: SlotCrossMoE+SlotRouter = CACS; FiLM feat*(1+sigmoid(gate)) = OCVA; num_slots=16, alpha=0.8, DINOv2-Base).
• Claims 2-5 (benchmark numbers): consistency-verified, not retrained. Every reported delta reproduces arithmetically from the paper's own tables (14/14 checks); a toy cross-view retrieval reproduces the direction of the Table-5 structural-loss ablation (+2.78 R@1). Full retraining (DINOv2-Base + Mixer, 448², 40 epochs on an 80GB A100 over University-1652 / SUES-200 / GTA-V) was not performed: no local GPU (Apple M1 Pro) and HF GPU Jobs return 402 Payment Required; no public checkpoint.

Links

• Paper: https://arxiv.org/abs/2605.07099 · OpenReview: https://openreview.net/forum?id=v6xrfEjNQv
• Official code: https://github.com/HRT00/Official_InfoGeo
• Baseline CVcities: https://github.com/GaoShuang98/CVCities · Backbone: https://huggingface.co/facebook/dinov2-base
• Datasets: https://huggingface.co/datasets/layumi/university-1652 · https://github.com/Reza-Zhu/SUES-200-Benchmark · https://huggingface.co/datasets/Yux1ang/GTA-UAV-LR · https://github.com/Dmmm1997/DenseUAV