pone7 commited on
Commit
91f5a2b
·
verified ·
1 Parent(s): ec02cc4

data: upload README.md

Browse files
Files changed (1) hide show
  1. README.md +130 -3
README.md CHANGED
@@ -1,3 +1,130 @@
1
- ---
2
- license: mit
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-4.0
3
+ pretty_name: AET-Bench
4
+ tags:
5
+ - atomic-electron-tomography
6
+ - haadf-stem
7
+ - 3d-reconstruction
8
+ - 4d-imaging
9
+ - materials-science
10
+ ---
11
+
12
+ # AET-Bench: A Comprehensive Benchmark for Atomic Electron Tomography
13
+
14
+ A NeurIPS 2026 Datasets & Benchmarks submission: 755 samples, 6 difficulty
15
+ tiers, 14 primary reconstruction methods, and 4 evaluation tasks spanning
16
+ static 3D, atom-level localization, element classification, and 4D dynamic
17
+ tracking of nanoparticles at the atomic scale.
18
+
19
+ AET-Bench introduces **atom identity persistence** and **trajectory RMSD**
20
+ — two atom-centric metrics that expose a striking failure mode:
21
+ state-of-the-art 4D Gaussian-splatting reconstructions reach SSIM 0.94 on
22
+ AET data yet track only 15 % of individual atoms across time. The paper
23
+ argues that pixel-level metrics systematically mislead atomic-scale
24
+ evaluation because the HAADF-STEM $Z^\gamma$ forward model decouples density
25
+ from atom count.
26
+
27
+ ## Highlights
28
+
29
+ | | |
30
+ |---|---|
31
+ | **Samples** | 755 across 6 tiers (analytical, binary alloys, 4D dynamics, multislice, real, Brownian Pt) |
32
+ | **Methods** | 14 primary + supplementary baselines from 5 algorithmic families |
33
+ | **Tasks** | T1 static 3D · T2 atom localization · T3 element classification · T7 4D dynamics |
34
+ | **Results** | 14,000+ rows in `data/aet_bench/results.jsonl` |
35
+ | **Findings** | 7 findings including 4D paradox, dose sensitivity, sim-to-real gap |
36
+
37
+ ## Quickstart
38
+
39
+ ```bash
40
+ # Install
41
+ conda env create -f environment.yml
42
+ conda activate aet-bench
43
+
44
+ # Run one method on one sample
45
+ python scripts/run_benchmark.py \
46
+ --subset mono --sample Cu-0165-00 \
47
+ --method wbp --task T1
48
+
49
+ # Regenerate paper figures & tables
50
+ python scripts/generate_paper_assets.py
51
+
52
+ # Audit the paper result log and record its SHA256/provenance
53
+ python scripts/validate_results.py data/aet_bench/results.jsonl
54
+ python scripts/generate_paper_assets.py --results data/aet_bench/results.jsonl
55
+
56
+ # Fill GPU-heavy missing baselines; defaults to one worker per visible GPU
57
+ # Uses the existing aet4dgs conda env unless PYTHON_BIN is set.
58
+ bash scripts/start_gpu_extras.sh
59
+
60
+ # 4D-paradox mechanism analysis (reproduces Appendix D)
61
+ python scripts/analysis/fourd_mechanism.py
62
+ ```
63
+
64
+ ## Repository Layout
65
+
66
+ ```
67
+ data/aet_bench/ # benchmark data + results.jsonl (git-ignored)
68
+ scripts/
69
+ run_benchmark.py # unified entry point: --subset --method --task
70
+ generate_paper_assets.py # regenerates every paper figure + LaTeX table
71
+ bench_bootstrap.py # 95 % bootstrap CI utility
72
+ analysis/ # paradox / dose / ranking analyses
73
+ metrics/ # FSC, Hungarian atom matching, trajectory metrics
74
+ baselines/ # thin wrappers around each method
75
+ paper/aet_bench.tex # NeurIPS submission
76
+ paper/croissant.json # Croissant 1.0 dataset card
77
+ ```
78
+
79
+ ## Adding a New Method
80
+
81
+ Methods are registered in `scripts/run_benchmark.py` under `METHOD_REGISTRY`.
82
+ Each entry maps a method name to a callable `f(data, sample_id) -> volume`.
83
+ See `CONTRIBUTING.md` for the detailed contract and examples.
84
+
85
+ ## Data
86
+
87
+ Simulated tiers are reproducible from scratch via
88
+ `scripts/generate_haadf_data.py` (analytical $Z^\gamma$),
89
+ `scripts/generate_abtem_multislice.py` (dynamical scattering), and
90
+ `scripts/generate_dose_tier.py` (dose-controlled Poisson noise).
91
+
92
+ Real experimental tier sources 26 nanoparticles from published Zenodo
93
+ records (DOIs listed in `paper/croissant.json`). The NeurIPS review
94
+ release bundle is hosted at
95
+ `https://huggingface.co/datasets/pone7/4D-ATE`.
96
+
97
+ ## Citation
98
+
99
+ Submitted to NeurIPS 2026 Datasets & Benchmarks. BibTeX will be added
100
+ after the review window.
101
+
102
+ ## License
103
+
104
+ CC-BY-4.0 for the benchmark data and paper; MIT for the reference
105
+ reconstruction and evaluation code.
106
+
107
+ ## Hugging Face Large-File Parts
108
+
109
+ Some tier archives are split into parts to keep individual Hub files below the recommended size limit.
110
+ Concatenate parts in lexical order before extracting the reconstructed tar archive:
111
+
112
+ ```bash
113
+ cat aet-bench-bi.tar.part-* > aet-bench-bi.tar
114
+ sha256sum aet-bench-bi.tar
115
+ ```
116
+ Expected SHA256 for `aet-bench-bi.tar`: `6661b6b85c267e28010d479e854a95a84cf5bab8dde45cc0ed6a247b7a85b3e2`.
117
+
118
+ ```bash
119
+ cat aet-bench-mono.tar.part-* > aet-bench-mono.tar
120
+ sha256sum aet-bench-mono.tar
121
+ ```
122
+ Expected SHA256 for `aet-bench-mono.tar`: `926adcfb2fc73c1d62b8c44befc9963540899f55a07669022da99c0376674602`.
123
+
124
+ ```bash
125
+ cat aet-bench-real.tar.part-* > aet-bench-real.tar
126
+ sha256sum aet-bench-real.tar
127
+ ```
128
+ Expected SHA256 for `aet-bench-real.tar`: `7e433949b0c53d739b594bf724d4505e7e348b7ca1113e0838da630ae242e43d`.
129
+
130
+ Uploaded-file checksums are in `SHA256SUMS`; reconstructed archive checksums are in `ORIGINAL_ARCHIVE_SHA256SUMS`.