ympradyumna commited on
Commit
8eb4033
·
verified ·
1 Parent(s): 2d826a7

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +99 -0
README.md ADDED
@@ -0,0 +1,99 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ # NOTE: pick a licence before making this repo public -- see "Licence" below.
3
+ license: other
4
+ license_name: see-licence-section
5
+ pretty_name: HMR scene-aware evaluation release
6
+ size_categories:
7
+ - n<1K
8
+ tags:
9
+ - 3d-human-pose
10
+ - human-scene-interaction
11
+ - smpl-x
12
+ - contact-estimation
13
+ ---
14
+
15
+ # HMR scene-aware evaluation release
16
+
17
+ Per-frame predictions for reproducing our quantitative results on
18
+ **PROX-quantitative** (178 frames) and **RICH** (100 frames), together with the
19
+ reconstructed scene each one was optimised against.
20
+
21
+ The evaluation code lives in the `eval_release/` folder of the code repository;
22
+ this dataset holds only the data it reads.
23
+
24
+ ## What is here
25
+
26
+ ```
27
+ release_data_full/prox/<recording>/<frame>.npz full scene clouds, 7.6 GB
28
+ release_data_full/rich/<frame>.npz 5.6 GB
29
+ release_data/prox/<recording>/<frame>.npz compact tier, 15 MB
30
+ release_data/rich/<frame>.npz 9.6 MB
31
+ ```
32
+
33
+ Both tiers hold the same predictions and differ only in how much of the
34
+ reconstructed scene comes with them:
35
+
36
+ * **`release_data_full/`** -- the complete point cloud the optimiser saw, ~2.0M
37
+ points per PROX frame and ~2.6M per RICH frame, floor plane included. This is
38
+ the reference: nothing about the scene is withheld, so the contact numbers can
39
+ be checked against the whole reconstruction.
40
+ * **`release_data/`** -- the same predictions with the scene reduced to the
41
+ points the contact metric can reach (~3.6k and ~4.5k points per frame). Small
42
+ enough to sit next to the code.
43
+
44
+ Evaluating either gives the same answer. Across all 278 frames every pose and
45
+ contact metric agrees exactly; only the vertex-to-scene *vector* metric moves,
46
+ by <1e-7 relative, where two scene points are equidistant from a body vertex to
47
+ float32 precision and a nearest-neighbour search picks between them arbitrarily.
48
+
49
+ ## Per-frame fields
50
+
51
+ | field | what it is |
52
+ |---|---|
53
+ | `init_*`, `final_*` | SMPL-X parameters (betas + 6D rotations) and camera translation, before and after scene-aware optimisation |
54
+ | `K` | predicted pinhole intrinsics |
55
+ | `contact_deco` | the image-based contact prediction the optimiser was given |
56
+ | `scene_points`, `scene_normals` | reconstructed scene, metric scale, camera frame |
57
+ | `visibility_mask` | body vertices projecting into the segmented person (PROX only) |
58
+ | `contact_pred_reference` | contact mask as computed by the pipeline, for verification |
59
+
60
+ ## What is *not* here
61
+
62
+ No dataset imagery, depth maps, meshes or ground truth. The scene points are
63
+ geometry our pipeline reconstructed from the input frame, not pixels. Ground
64
+ truth comes from your own PROX and RICH downloads, used as distributed:
65
+
66
+ * PROX -- https://prox.is.tue.mpg.de
67
+ * RICH -- https://rich.is.tue.mpg.de
68
+ * SMPL-X -- https://smpl-x.is.tue.mpg.de
69
+
70
+ Each is under its own licence, which you accept with the dataset provider.
71
+
72
+ ## Usage
73
+
74
+ ```python
75
+ from huggingface_hub import snapshot_download
76
+
77
+ # compact tier only -- enough to reproduce every number
78
+ snapshot_download('ympradyumna/hmr-eval-release', repo_type='dataset',
79
+ local_dir='eval_release',
80
+ allow_patterns='release_data/*')
81
+
82
+ # or the full scene clouds (~13 GB)
83
+ snapshot_download('ympradyumna/hmr-eval-release', repo_type='dataset',
84
+ local_dir='eval_release',
85
+ allow_patterns='release_data_full/*')
86
+ ```
87
+
88
+ Then, from the code repository:
89
+
90
+ ```bash
91
+ python eval_release/eval_prox.py --prox-root /path/to/PROX/quantitative --smplx-dir /path/to/smplx
92
+ python eval_release/eval_rich.py --rich-root /path/to/rich --data-root /path/to/data
93
+ ```
94
+
95
+ ## Licence
96
+
97
+ **To be decided before this repo is made public.** The predictions are ours to
98
+ license, but they are derived from PROX and RICH inputs, so the terms should be
99
+ compatible with those datasets' licences.