zeredata-admin commited on
Commit
fbfb41a
·
verified ·
1 Parent(s): b146fda

Add files using upload-large-folder tool

Browse files
CITATION.bib CHANGED
@@ -1,7 +1,7 @@
1
- @dataset{zeredata_binpicking_2026,
2
- author = {Umit Kavala},
3
- title = {ZereData Bin Picking Dataset v1.1},
4
- year = {2026},
5
- publisher = {HuggingFace},
6
- url = {https://huggingface.co/datasets/zeredata/bin-picking-v1}
7
- }
 
1
+ @dataset{zeredata_binpicking_2026,
2
+ author = {Umit Kavala},
3
+ title = {ZereData Bin Picking Dataset v1.2},
4
+ year = {2026},
5
+ publisher = {HuggingFace},
6
+ url = {https://huggingface.co/datasets/zeredata/bin-picking}
7
+ }
LICENSE CHANGED
@@ -1,22 +1,22 @@
1
- Attribution 4.0 International (CC BY 4.0)
2
-
3
- By exercising the Licensed Rights (defined below), You accept and agree to be
4
- bound by the terms and conditions of this Creative Commons Attribution 4.0
5
- International Public License ("Public License"). To the extent this Public
6
- License may be interpreted as a contract, You are granted the Licensed Rights
7
- in consideration of Your acceptance of these terms and conditions, and the
8
- Licensor grants You such rights in consideration of benefits the Licensor
9
- receives from making the Licensed Material available under these terms and
10
- conditions.
11
-
12
- The full legal text is available at:
13
- https://creativecommons.org/licenses/by/4.0/legalcode
14
-
15
- Summary of rights:
16
- - Share — copy and redistribute the material in any medium or format.
17
- - Adapt — remix, transform, and build upon the material for any purpose,
18
- even commercially.
19
- - Attribution — You must give appropriate credit, provide a link to the
20
- license, and indicate if changes were made.
21
-
22
- No warranties are given.
 
1
+ Attribution 4.0 International (CC BY 4.0)
2
+
3
+ By exercising the Licensed Rights (defined below), You accept and agree to be
4
+ bound by the terms and conditions of this Creative Commons Attribution 4.0
5
+ International Public License ("Public License"). To the extent this Public
6
+ License may be interpreted as a contract, You are granted the Licensed Rights
7
+ in consideration of Your acceptance of these terms and conditions, and the
8
+ Licensor grants You such rights in consideration of benefits the Licensor
9
+ receives from making the Licensed Material available under these terms and
10
+ conditions.
11
+
12
+ The full legal text is available at:
13
+ https://creativecommons.org/licenses/by/4.0/legalcode
14
+
15
+ Summary of rights:
16
+ - Share — copy and redistribute the material in any medium or format.
17
+ - Adapt — remix, transform, and build upon the material for any purpose,
18
+ even commercially.
19
+ - Attribution — You must give appropriate credit, provide a link to the
20
+ license, and indicate if changes were made.
21
+
22
+ No warranties are given.
README.md CHANGED
@@ -1,5 +1,5 @@
1
  ---
2
- pretty_name: ZereData Bin Picking Dataset v1.1
3
  license: cc-by-4.0
4
  task_categories:
5
  - object-detection
@@ -24,9 +24,10 @@ language:
24
  - en
25
  ---
26
 
27
- # ZereData Bin Picking Dataset v1.1
28
 
29
- ![license](https://img.shields.io/badge/license-CC--BY--4.0-blue) ![scenes](https://img.shields.io/badge/scenes-10,000-green) ![size](https://img.shields.io/badge/size-14.8GB-orange)
 
 
30
 
31
  Synthetic training data for robotic bin picking — RGB, depth, instance masks, 6D pose, 2D bounding boxes, and per-instance visibility, in BOP/COCO/YOLO formats.
32
 
@@ -34,10 +35,12 @@ Synthetic training data for robotic bin picking — RGB, depth, instance masks,
34
 
35
  ## Overview
36
 
37
- Generated with GPU-accelerated, physically based ray tracing, this dataset delivers dense, photorealistic scenes of cluttered bins at warehouse scale. Each scene includes RGB, 32-bit depth, instance segmentation, camera intrinsics/extrinsics, and per-instance 6D pose with visibility ratios.
38
 
39
  The dataset's value is simple: synthetic renders give perfect ground truth annotations impossible to obtain from real cameras, at a scale and cost real-world collection cannot match. Use it to train 6D pose estimators, bin-picking grasp predictors, and warehouse perception systems — then validate sim-to-real transfer on smaller real-world test sets.
40
 
 
 
41
  ## Dataset Statistics
42
 
43
  | Metric | Value |
@@ -46,24 +49,26 @@ The dataset's value is simple: synthetic renders give perfect ground truth annot
46
  | Train split | 8,000 |
47
  | Val split | 2,000 |
48
  | Resolution | 1280x720 |
49
- | Object instances | 296,603 |
50
- | Object categories | 4 |
 
 
51
  | Modalities | 6 (RGB, depth, mask, pose, bboxes, visibility) |
52
- | Total size on disk | 14.8 GB |
53
 
54
  ## Modalities
55
 
56
  - **RGB** — 1280×720 PNG per scene. The primary input for detection, segmentation, and pose models.
57
- - **Depth** — 32-bit EXR in metres. Train depth-conditioned pose models or use as a second-channel input.
58
  - **Instance mask** — colour-coded PNG per scene, one colour per object instance. Drives instance segmentation and occlusion reasoning.
59
- - **6D pose** — per-instance rotation and translation in camera frame (BOP `cam_R_m2c`, `cam_t_m2c`). Supervises pose regression heads.
60
- - **2D bounding boxes** — derived from masks, included in COCO and YOLO formats.
61
- - **Visibility ratio** — BOP `visib_fract` per instance; lets you weight the training loss by occlusion severity.
62
 
63
  ## Formats
64
 
65
  ### BOP (primary)
66
- Canonical BOP directory layout under `data/train/` and `data/val/`. Each scene folder contains `scene_camera.json` (`cam_K`, `depth_scale`), `scene_gt.json` (per-object `cam_R_m2c`, `cam_t_m2c`, `obj_id`), and `scene_gt_info.json` (`bbox_obj`, `bbox_visib`, `visib_fract`). Load with the BOP toolkit. Object IDs are **ZereData-specific, not BOP canonical** — see Limitations.
67
 
68
  ### COCO
69
  Merged `annotations/coco_train.json` and `annotations/coco_val.json` with `images`, `annotations` (bboxes + masks), and `categories`. Loads cleanly with pycocotools:
@@ -76,6 +81,55 @@ coco = COCO('annotations/coco_train.json')
76
  ### YOLO
77
  Per-image `.txt` label files under `annotations/yolo_train/` and `yolo_val/`, with normalized `class_id cx cy w h` entries. Class IDs are consistent across both splits; see `annotations/yolo_classes.txt` and `annotations/yolo_data.yaml`.
78
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
79
  ## Data Format
80
 
81
  This dataset is packaged as per-format zip archives, mirroring the [bop-benchmark](https://huggingface.co/bop-benchmark) HF layout convention (one zip per logical split) adapted for multi-format shipping. Loose files — README, LICENSE, CITATION, metadata.json, preview images — remain at the repository root so the HF dataset page renders a preview.
@@ -87,7 +141,7 @@ This dataset is packaged as per-format zip archives, mirroring the [bop-benchmar
87
  | `bin_picking_coco.zip` | `coco_train.json`, `coco_val.json` (merged, BOP obj IDs remapped to COCO categories) | `annotations/coco_*.json` |
88
  | `bin_picking_yolo.zip` | YOLO labels per split + `yolo_classes.txt` + `yolo_data.yaml` | `annotations/yolo_{train,val}/*.txt`, `annotations/yolo_*.{txt,yaml}` |
89
  | `bin_picking_native.zip` | Per-scene native annotations (full pre-export ZereData scene graph) | `annotations/scene_NNNN.json` |
90
- | `bin_picking_models.zip` | 27 GLB object models | `models/*.glb` |
91
 
92
  ### Download and extract only what you need
93
 
@@ -95,7 +149,7 @@ This dataset is packaged as per-format zip archives, mirroring the [bop-benchmar
95
  from huggingface_hub import hf_hub_download
96
  import zipfile
97
 
98
- REPO = 'zeredata/bin-picking-v1'
99
  # BOP train split
100
  p = hf_hub_download(repo_id=REPO, filename='bin_picking_train_bop.zip', repo_type='dataset')
101
  with zipfile.ZipFile(p) as z:
@@ -105,7 +159,7 @@ with zipfile.ZipFile(p) as z:
105
  Or the whole dataset in one shot:
106
 
107
  ```bash
108
- huggingface-cli download --repo-type dataset zeredata/bin-picking-v1 --local-dir ./zd_bp
109
  cd ./zd_bp && for z in bin_picking_*.zip; do unzip -q "$z"; done
110
  ```
111
 
@@ -146,27 +200,23 @@ for ann in coco.loadAnns(coco.getAnnIds(imgIds=img_ids[0])):
146
  print(ann['bbox'], ann['category_id'])
147
  ```
148
 
149
- _A `datasets.load_dataset()` loader is planned for v1.1._
150
-
151
  ## Intended Use
152
 
153
  Training 6D pose estimation models, bin-picking grasp models, and warehouse robotics perception systems. Synthetic data for sim-to-real transfer research.
154
 
155
  ## Pose convention
156
 
157
- This release (v1.0 and v1.1) ships 6D poses in **OpenGL camera convention** (camera looks down `-Z`, in-front objects have `cam_t_m2c.z < 0`) rather than the BOP-standard **OpenCV convention** (camera looks down `+Z`, in-front objects have `cam_t_m2c.z > 0`). The two differ by the basis change `diag(1, -1, -1)` applied to the camera frame.
158
 
159
- **Anyone scoring against bop_toolkit_lib, MegaPose, FoundationPose, CosyPose, or any OpenCV-convention model must apply `diag(1, -1, -1)` to the GT `cam_R_m2c` and `cam_t_m2c` from `scene_gt.json` before evaluation**, otherwise pose errors blow up to the order of the object diameter (~150 mm) and any AR comparison is meaningless.
160
 
161
- ZereData's evaluation harness exposes this via the `--legacy-gl-convention` flag on `eval.scorer.BopScorer`, `eval.adapter.gt_as_predictions`, and the eval CLIs (`gt_sanity`, `run_megapose_eval`, `rescore`, `_perturb_probe`, `rescore_h1`). Pass it when scoring v1.0 or v1.1; omit it for v1.2 onward, which ships in OpenCV convention at the producer.
162
 
163
- This deviation was identified post-publication. v1.0 and v1.1 remain available unchanged on HuggingFace for reproducibility; **v1.2 supersedes them for new integrations** and ships BOP-spec compliant out of the box. The legacy flag is supported indefinitely so downstream code that already ingests v1.0/v1.1 keeps working.
164
-
165
- RGB, depth, masks, COCO 2D boxes, and YOLO labels are **unaffected** — only the 6D pose serialisation deviates from the BOP spec.
166
 
167
  ## Limitations and Known Issues
168
 
169
- - **BOP pose convention deviation.** Object pose extrinsics in `scene_gt.json` ship in OpenGL convention, not BOP-spec OpenCV. See the **Pose convention** section above for the basis change and the `--legacy-gl-convention` flag. v1.2 supersedes this with a producer-side fix.
170
  - **Warehouse-specific lighting.** The three lighting profiles model warehouse conditions and may not transfer directly to outdoor, medical, or agricultural domains:
171
  - `bin_picking_overhead` — bright fluorescent overhead panels, typical of distribution-center shelving aisles.
172
  - `bin_picking_mixed` — mixed overhead + rim lighting with warmer colour temperature, mimicking older facilities with partial skylights.
@@ -175,21 +225,17 @@ RGB, depth, masks, COCO 2D boxes, and YOLO labels are **unaffected** — only th
175
  - **Procedural materials.** Material variation uses procedural textures, not photoscanned assets. High-frequency surface detail may look synthetic under close inspection.
176
  - **Geometric occlusion only.** No category-level occlusion modelling — occlusion is derived from geometry alone.
177
  - **Simulated camera intrinsics.** The intrinsic matrix is synthetic, not drawn from real sensor calibration.
178
-
179
- ## Evaluation
180
-
181
- Benchmark evaluation on LM-O is forthcoming; see [ZereData](https://zeredata.com) for updates.
182
 
183
  ## Comparison to Related Datasets
184
 
185
  HOPE, T-LESS, and YCB-Video are excellent real-world datasets with limited scale and fixed object sets. This dataset is synthetic-only, scales without bound, and supports customer-specific object libraries. Treat the two as complementary: real data for evaluation, synthetic data for training.
186
 
187
-
188
  ## Custom Datasets
189
 
190
  This release is a research dataset. The categories (bottle, box, can, pouch), SKU shapes, and bin geometry are intentionally generic — useful for benchmarking, pretraining, and sanity-checking a 6D pose pipeline before you invest in real-world data collection.
191
 
192
- For production use, ZereData generates the same kind of dataset matched to your warehouse's actual SKUs and bin geometry. Customer-specific datasets ingest CAD files or reference photos, render at the same scale and quality as this release, and ship in days. Pricing is per-dataset, with design-partner terms for early customers.
193
 
194
  If you're training bin-picking models for a specific picking environment, email **engineering@zeredata.com** — design partners welcome.
195
 
@@ -198,7 +244,7 @@ If you're training bin-picking models for a specific picking environment, email
198
  ```bibtex
199
  @dataset{zeredata_binpicking_2026,
200
  author = {Umit Kavala},
201
- title = {ZereData Bin Picking Dataset v1.1},
202
  year = {2026},
203
  publisher = {HuggingFace},
204
  url = {https://huggingface.co/datasets/zeredata/bin-picking}
 
1
  ---
2
+ pretty_name: ZereData Bin Picking Dataset v1.2
3
  license: cc-by-4.0
4
  task_categories:
5
  - object-detection
 
24
  - en
25
  ---
26
 
 
27
 
28
+ # ZereData Bin Picking Dataset v1.2
29
+
30
+ ![license](https://img.shields.io/badge/license-CC--BY--4.0-blue) ![scenes](https://img.shields.io/badge/scenes-10,000-green) ![size](https://img.shields.io/badge/size-8.8GB-orange)
31
 
32
  Synthetic training data for robotic bin picking — RGB, depth, instance masks, 6D pose, 2D bounding boxes, and per-instance visibility, in BOP/COCO/YOLO formats.
33
 
 
35
 
36
  ## Overview
37
 
38
+ Generated with GPU-accelerated, physically based ray tracing, this dataset delivers photorealistic scenes of cluttered bins at warehouse scale. Each scene includes RGB, depth, instance segmentation, camera intrinsics/extrinsics, and per-instance 6D pose with visibility ratios.
39
 
40
  The dataset's value is simple: synthetic renders give perfect ground truth annotations impossible to obtain from real cameras, at a scale and cost real-world collection cannot match. Use it to train 6D pose estimators, bin-picking grasp predictors, and warehouse perception systems — then validate sim-to-real transfer on smaller real-world test sets.
41
 
42
+ **v1.2 is the recommended version for new integrations.** It ships BOP-spec pose convention at the producer, real per-instance visibility, object models with symmetry annotations, and corrected object placement. See **Versions** below for what changed and why v1.0/v1.1 remain available.
43
+
44
  ## Dataset Statistics
45
 
46
  | Metric | Value |
 
49
  | Train split | 8,000 |
50
  | Val split | 2,000 |
51
  | Resolution | 1280x720 |
52
+ | Object instances | 162,897 |
53
+ | Object instances per scene | 16.29 ± 10.23 |
54
+ | Object categories | 4 (bottle, box, can, pouch) |
55
+ | Distinct object models (SKUs) | 24 |
56
  | Modalities | 6 (RGB, depth, mask, pose, bboxes, visibility) |
57
+ | Total size on disk | 8.8 GB |
58
 
59
  ## Modalities
60
 
61
  - **RGB** — 1280×720 PNG per scene. The primary input for detection, segmentation, and pose models.
62
+ - **Depth** — 16-bit (uint16) single-channel PNG in millimetres, read as `mm = pixel * depth_scale` using the per-scene `depth_scale` in `scene_camera.json`; a pixel value of 0 means "no measurement". Train depth-conditioned pose models or use as a second-channel input.
63
  - **Instance mask** — colour-coded PNG per scene, one colour per object instance. Drives instance segmentation and occlusion reasoning.
64
+ - **6D pose** — per-instance rotation and translation in camera frame (BOP `cam_R_m2c`, `cam_t_m2c`), in **OpenCV convention** as the BOP spec requires. Supervises pose regression heads.
65
+ - **2D bounding boxes** — amodal (`bbox_obj`) and visible (`bbox_visib`), included in COCO and YOLO formats.
66
+ - **Visibility ratio** — BOP `visib_fract` per instance, computed from amodal vs visible pixel counts; lets you weight the training loss by occlusion severity.
67
 
68
  ## Formats
69
 
70
  ### BOP (primary)
71
+ Canonical BOP directory layout under `data/train/` and `data/val/`. Each scene folder contains `scene_camera.json` (`cam_K`, `depth_scale`), `scene_gt.json` (per-object `cam_R_m2c`, `cam_t_m2c`, `obj_id`), and `scene_gt_info.json` (`bbox_obj`, `bbox_visib`, `visib_fract`). `models/` ships the 24 object meshes as `obj_NNNNNN.ply` — load these with the BOP toolkit — alongside `models_info.json` (mm units, computed from the PLY geometry, including per-model symmetry annotations). The same 24 models also ship as `.glb` for visualization/CAD use, indexed by `models_index.json`; both the PLY pack and the GLB index are keyed by the same `obj_id` as `scene_gt.json`. Object IDs are **ZereData-specific, not BOP canonical** — see Limitations.
72
 
73
  ### COCO
74
  Merged `annotations/coco_train.json` and `annotations/coco_val.json` with `images`, `annotations` (bboxes + masks), and `categories`. Loads cleanly with pycocotools:
 
81
  ### YOLO
82
  Per-image `.txt` label files under `annotations/yolo_train/` and `yolo_val/`, with normalized `class_id cx cy w h` entries. Class IDs are consistent across both splits; see `annotations/yolo_classes.txt` and `annotations/yolo_data.yaml`.
83
 
84
+ ## Versions
85
+
86
+ | Version | Status | Pose convention | Notes |
87
+ |---|---|---|---|
88
+ | v1.0 | Superseded, retained | OpenGL (see below) | First public release. |
89
+ | v1.1 | Superseded, retained | OpenGL (see below) | Packaging and metadata pass over v1.0. |
90
+ | **v1.2** | **Recommended** | **OpenCV (BOP spec)** | Producer-side convention fix, real `visib_fract`, amodal + visible boxes, symmetry annotations, object models shipped, corrected placement geometry. |
91
+
92
+ v1.0 and v1.1 remain available unchanged for reproducibility. v1.2 is not a drop-in
93
+ replacement for a pipeline already calibrated against v1.1: the pose convention differs,
94
+ and the scene composition differs (see **Dataset characteristics** below).
95
+
96
+ ### Generation stack per version
97
+
98
+ v1.0 and v1.1 were generated on the previous production rendering stack; v1.2 is
99
+ generated on the current one. Both stacks use the same GPU path tracer and the same
100
+ scene-generation code path. Renders are reproducible at the **dataset level** — same
101
+ scene count, same modalities, same export formats, same configuration — and are **not
102
+ bit-reproducible**: re-running a scene with the same seed produces a visually equivalent
103
+ but not pixel-identical image. This is a property of the pipeline and is true within a
104
+ single stack as well as across stacks. If you need bit-exact inputs, pin the published
105
+ archives rather than regenerating.
106
+
107
+ ## Dataset characteristics
108
+
109
+ These are properties of how the dataset is composed, not defects. They matter if you are
110
+ sizing a training run for a particular class.
111
+
112
+ - **Class mix is not balanced, and box instances are sparse.** Scenes are drawn
113
+ from two object-size groups: a small-object group that admits up to 33 objects per
114
+ scene, and a large-object group capped at 7. All six box models fall in the large-object
115
+ group, so boxes appear in roughly half the scenes and only a few at a time. Boxes are
116
+ **9.1% of all object instances** in v1.2, against roughly 26% in v1.1 —
117
+ about one third the box instances per scene. Bottles, cans and pouches carry the
118
+ remainder. **If you are training a box detector, size your run against the box instance
119
+ count, not the scene count.** This composition is deliberate and is not tuned toward a
120
+ target class balance; a production dataset is built around the customer's own SKUs and
121
+ bin geometry, which sets the mix directly.
122
+ - **A small number of scenes contain no box instances despite drawing from the
123
+ box-bearing group.** Where a scene draws a single box and that box does not fit the
124
+ remaining space, the scene ships without it. Measured at 0.4% (36 of 10,000) of
125
+ scenes. Scenes are individually correct — the annotation matches the image — but a
126
+ per-scene guarantee of box presence is not one of them.
127
+ - **Objects per scene varies widely by size group.** Small-object scenes are dense; the
128
+ large-object scenes are sparse by construction. The per-scene instance count in the
129
+ statistics table is a mean across both.
130
+ - **One pouch model.** The pouch class is represented by a single mesh, so pouch
131
+ appearance variation comes from pose, lighting and materials rather than shape.
132
+
133
  ## Data Format
134
 
135
  This dataset is packaged as per-format zip archives, mirroring the [bop-benchmark](https://huggingface.co/bop-benchmark) HF layout convention (one zip per logical split) adapted for multi-format shipping. Loose files — README, LICENSE, CITATION, metadata.json, preview images — remain at the repository root so the HF dataset page renders a preview.
 
141
  | `bin_picking_coco.zip` | `coco_train.json`, `coco_val.json` (merged, BOP obj IDs remapped to COCO categories) | `annotations/coco_*.json` |
142
  | `bin_picking_yolo.zip` | YOLO labels per split + `yolo_classes.txt` + `yolo_data.yaml` | `annotations/yolo_{train,val}/*.txt`, `annotations/yolo_*.{txt,yaml}` |
143
  | `bin_picking_native.zip` | Per-scene native annotations (full pre-export ZereData scene graph) | `annotations/scene_NNNN.json` |
144
+ | `bin_picking_models.zip` | 24 BOP-toolkit PLY meshes + 24 GLB object models + `models_info.json` (from the PLY geometry) + `models_index.json` (obj_id → GLB filename) | `models/*.ply`, `models/*.glb`, `models/models_info.json`, `models/models_index.json` |
145
 
146
  ### Download and extract only what you need
147
 
 
149
  from huggingface_hub import hf_hub_download
150
  import zipfile
151
 
152
+ REPO = 'zeredata/bin-picking'
153
  # BOP train split
154
  p = hf_hub_download(repo_id=REPO, filename='bin_picking_train_bop.zip', repo_type='dataset')
155
  with zipfile.ZipFile(p) as z:
 
159
  Or the whole dataset in one shot:
160
 
161
  ```bash
162
+ huggingface-cli download --repo-type dataset zeredata/bin-picking --local-dir ./zd_bp
163
  cd ./zd_bp && for z in bin_picking_*.zip; do unzip -q "$z"; done
164
  ```
165
 
 
200
  print(ann['bbox'], ann['category_id'])
201
  ```
202
 
 
 
203
  ## Intended Use
204
 
205
  Training 6D pose estimation models, bin-picking grasp models, and warehouse robotics perception systems. Synthetic data for sim-to-real transfer research.
206
 
207
  ## Pose convention
208
 
209
+ **v1.2 ships 6D poses in OpenCV camera convention** (camera looks down `+Z`, in-front objects have `cam_t_m2c.z > 0`), as the BOP spec requires. No transform is needed to score against bop_toolkit_lib, MegaPose, FoundationPose, CosyPose, or any other OpenCV-convention model.
210
 
211
+ v1.0 and v1.1 shipped poses in **OpenGL** convention (camera looks down `-Z`, in-front objects have `cam_t_m2c.z < 0`). The two differ by the basis change `diag(1, -1, -1)` applied to the camera frame. **Anyone scoring v1.0 or v1.1 must apply `diag(1, -1, -1)` to the GT `cam_R_m2c` and `cam_t_m2c` before evaluation**, otherwise pose errors blow up to the order of the object diameter and any comparison is meaningless.
212
 
213
+ ZereData's evaluation harness exposes this via the `--legacy-gl-convention` flag. Pass it when scoring v1.0 or v1.1; **omit it for v1.2**. The legacy flag is supported indefinitely so downstream code that already ingests v1.0/v1.1 keeps working.
214
 
215
+ RGB, depth, masks, COCO 2D boxes, and YOLO labels are **unaffected** in every version — only the 6D pose serialisation deviated.
 
 
216
 
217
  ## Limitations and Known Issues
218
 
219
+ - **Object IDs are ZereData-specific.** `obj_id` values index this dataset's own 24-model library and do not correspond to any BOP canonical object set. `models_info.json`, `models/*.ply`, `models/models_index.json`, and `models/*.glb` are all keyed by the same `obj_id` as `scene_gt.json` — `models/` is the authority for what each ID is.
220
  - **Warehouse-specific lighting.** The three lighting profiles model warehouse conditions and may not transfer directly to outdoor, medical, or agricultural domains:
221
  - `bin_picking_overhead` — bright fluorescent overhead panels, typical of distribution-center shelving aisles.
222
  - `bin_picking_mixed` — mixed overhead + rim lighting with warmer colour temperature, mimicking older facilities with partial skylights.
 
225
  - **Procedural materials.** Material variation uses procedural textures, not photoscanned assets. High-frequency surface detail may look synthetic under close inspection.
226
  - **Geometric occlusion only.** No category-level occlusion modelling — occlusion is derived from geometry alone.
227
  - **Simulated camera intrinsics.** The intrinsic matrix is synthetic, not drawn from real sensor calibration.
228
+ - **Class mix and per-scene box presence.** See **Dataset characteristics** above.
 
 
 
229
 
230
  ## Comparison to Related Datasets
231
 
232
  HOPE, T-LESS, and YCB-Video are excellent real-world datasets with limited scale and fixed object sets. This dataset is synthetic-only, scales without bound, and supports customer-specific object libraries. Treat the two as complementary: real data for evaluation, synthetic data for training.
233
 
 
234
  ## Custom Datasets
235
 
236
  This release is a research dataset. The categories (bottle, box, can, pouch), SKU shapes, and bin geometry are intentionally generic — useful for benchmarking, pretraining, and sanity-checking a 6D pose pipeline before you invest in real-world data collection.
237
 
238
+ For production use, ZereData generates the same kind of dataset matched to your warehouse's actual SKUs and bin geometry. Customer-specific datasets ingest CAD files or reference photos and render at the same scale and quality as this release.
239
 
240
  If you're training bin-picking models for a specific picking environment, email **engineering@zeredata.com** — design partners welcome.
241
 
 
244
  ```bibtex
245
  @dataset{zeredata_binpicking_2026,
246
  author = {Umit Kavala},
247
+ title = {ZereData Bin Picking Dataset v1.2},
248
  year = {2026},
249
  publisher = {HuggingFace},
250
  url = {https://huggingface.co/datasets/zeredata/bin-picking}
bin_picking_coco.zip CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:29fcf548fac65434728c92af400ddc8f3701f8f78a973c346d2df80986b2e636
3
- size 5591483
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:31630a9020206ef5a7dcc6b3790cb0d5eab6ff206d036da60931076a69ec39a7
3
+ size 3091646
bin_picking_models.zip CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:217db39452f4ca2730a329a289b3d2020349e0f84740c09756507ea5a7629ae9
3
- size 19094258
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bab16b831cb199ac98f6033f4cfe8e4f90bd31cb0d49c1a1b8d5b44bc07f6700
3
+ size 19021893
bin_picking_native.zip CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:82c236b47b4e336560ff41bff0ba6a6bb9e8eb111c0204c598db326bdec5e86f
3
- size 78502371
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c182cb333a93a50a5df1482c2ac569b83f84b0cfe49b0042452be5130fd355d6
3
+ size 58264806
bin_picking_train_bop.zip CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b32b9f0c40a6d875e4d9cce5c2f6b9f15b5a242dcf655e9ab1f41305fd0b0b63
3
- size 12469497512
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e8cd2d21bf356b015c7d0fcea230206a089dc985cc5c3c871cd1901f05b3b681
3
+ size 6995245823
bin_picking_val_bop.zip CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:53baeed1b34d1f8ab4637a06b6fa7b2cb4dfc07e722bd47a757a7e9d6311040b
3
- size 3124316766
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fd7a5eb5142a9aa3a104895557253ad5827bdbfd5de49d147d8ffbae88da08f5
3
+ size 1747090398
bin_picking_yolo.zip CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:afa686eb536c5a267145e05db31b2044e51312b5fc0cc1544504424e85ad629b
3
- size 5805393
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e3f3f0b802b7d8eecf86faba4725a9dd515eaa1b0e9cfd04952a08c1a0bb1619
3
+ size 3919498
metadata.json CHANGED
@@ -1,113 +1,110 @@
1
- {
2
- "name": "bin-picking-v1",
3
- "version": "1.1.0",
4
- "author": "Umit Kavala",
5
- "release_date": "2026-04-29T19:03:06.828982+00:00",
6
- "num_scenes_total": 10000,
7
- "num_scenes_train": 8000,
8
- "num_scenes_val": 2000,
9
- "resolution": [
10
- 1280,
11
- 720
12
- ],
13
- "generation": {
14
- "engine": "Blender Cycles",
15
- "blender_version": "5.0.1",
16
- "cycles_samples": 128,
17
- "render_device": "GPU (CUDA)",
18
- "avg_render_time_per_scene_seconds": 14.725
19
- },
20
- "objects": {
21
- "count": 27,
22
- "categories": {
23
- "bottles": [
24
- "beer_bottle_low_poly",
25
- "bottle",
26
- "bottle (1)",
27
- "bottle (2)",
28
- "coca_cola_bottle",
29
- "glass_soda_bottle",
30
- "liquid_detergent_bottle",
31
- "scooberts_beer_bottle",
32
- "simple_plastic_bottle",
33
- "water_bottle",
34
- "water_bottle_low_poly",
35
- "water_bottle_prop"
36
- ],
37
- "boxes": [
38
- "Box",
39
- "Cardboard Boxes",
40
- "Crate",
41
- "Fruit Crate",
42
- "cc0_free_cardboard_box",
43
- "low_poly_wooden_crate",
44
- "the_brown_cardboard_box"
45
- ],
46
- "cans": [
47
- "Gas Can",
48
- "metal_can",
49
- "soda_can",
50
- "soda_can (1)",
51
- "tin_can"
52
- ],
53
- "pouches": [
54
- "Bags",
55
- "Coin Pouch",
56
- "Sack Trench Small"
57
- ]
58
- },
59
- "object_id_scheme": "zeredata_v1",
60
- "bop_canonical_compatible": true
61
- },
62
- "category_map": {
63
- "bottle": 1,
64
- "box": 2,
65
- "can": 3,
66
- "pouch": 4
67
- },
68
- "scene_parameters": {
69
- "objects_per_scene": [
70
- 20,
71
- 40
72
- ],
73
- "bin_size_range_meters": [
74
- [
75
- 0.3,
76
- 0.3,
77
- 0.2
78
- ],
79
- [
80
- 0.6,
81
- 0.6,
82
- 0.4
83
- ]
84
- ],
85
- "lighting_profiles": [
86
- "bin_picking_overhead",
87
- "bin_picking_mixed",
88
- "studio"
89
- ],
90
- "bin_conditions": [
91
- "new",
92
- "scratched",
93
- "damaged"
94
- ]
95
- },
96
- "modalities": [
97
- "rgb",
98
- "depth",
99
- "instance_mask",
100
- "camera_intrinsics",
101
- "camera_extrinsics",
102
- "6d_pose",
103
- "bbox_2d",
104
- "visibility"
105
- ],
106
- "annotation_formats": [
107
- "bop",
108
- "coco",
109
- "yolo"
110
- ],
111
- "random_seed": 42,
112
- "generation_manifest_hash": "92bdf617d3e2e99ca466e6aef4724b27a94673afbe9bb75f18c53fe79ce23f27"
113
  }
 
1
+ {
2
+ "name": "bin-picking",
3
+ "version": "1.2.0",
4
+ "author": "Umit Kavala",
5
+ "release_date": "2026-09-06T17:01:51.890009+00:00",
6
+ "num_scenes_total": 10000,
7
+ "num_scenes_train": 8000,
8
+ "num_scenes_val": 2000,
9
+ "resolution": [
10
+ 1280,
11
+ 720
12
+ ],
13
+ "generation": {
14
+ "engine": "Blender Cycles",
15
+ "blender_version": "5.1.1",
16
+ "cycles_samples": 128,
17
+ "render_device": "GPU (CUDA, NVIDIA L4)",
18
+ "avg_render_time_per_scene_seconds": 0.0
19
+ },
20
+ "objects": {
21
+ "count": 24,
22
+ "categories": {
23
+ "bottles": [
24
+ "beer_bottle_low_poly",
25
+ "bottle (1)",
26
+ "bottle (2)",
27
+ "bottle",
28
+ "coca_cola_bottle",
29
+ "glass_soda_bottle",
30
+ "liquid_detergent_bottle",
31
+ "scooberts_beer_bottle",
32
+ "simple_plastic_bottle",
33
+ "water_bottle",
34
+ "water_bottle_low_poly",
35
+ "water_bottle_prop"
36
+ ],
37
+ "boxes": [
38
+ "Box",
39
+ "Crate",
40
+ "Fruit Crate",
41
+ "cc0_free_cardboard_box",
42
+ "low_poly_wooden_crate",
43
+ "the_brown_cardboard_box"
44
+ ],
45
+ "cans": [
46
+ "Gas Can",
47
+ "metal_can",
48
+ "soda_can (1)",
49
+ "soda_can",
50
+ "tin_can"
51
+ ],
52
+ "pouches": [
53
+ "Coin Pouch"
54
+ ]
55
+ },
56
+ "object_id_scheme": "zeredata_v1",
57
+ "bop_canonical_compatible": true
58
+ },
59
+ "category_map": {
60
+ "bottle": 1,
61
+ "box": 2,
62
+ "can": 3,
63
+ "pouch": 4
64
+ },
65
+ "scene_parameters": {
66
+ "objects_per_scene": [
67
+ 20,
68
+ 40
69
+ ],
70
+ "bin_size_range_meters": [
71
+ [
72
+ 0.3,
73
+ 0.3,
74
+ 0.2
75
+ ],
76
+ [
77
+ 0.6,
78
+ 0.6,
79
+ 0.4
80
+ ]
81
+ ],
82
+ "lighting_profiles": [
83
+ "bin_picking_overhead",
84
+ "bin_picking_mixed",
85
+ "studio"
86
+ ],
87
+ "bin_conditions": [
88
+ "new",
89
+ "scratched",
90
+ "damaged"
91
+ ]
92
+ },
93
+ "modalities": [
94
+ "rgb",
95
+ "depth",
96
+ "instance_mask",
97
+ "camera_intrinsics",
98
+ "camera_extrinsics",
99
+ "6d_pose",
100
+ "bbox_2d",
101
+ "visibility"
102
+ ],
103
+ "annotation_formats": [
104
+ "bop",
105
+ "coco",
106
+ "yolo"
107
+ ],
108
+ "random_seed": 42,
109
+ "generation_manifest_hash": "db6cfe6eaa0d348588c020941b742e3ccc09a76729706be0d26a20436a7605f7"
 
 
 
110
  }
preview/0001_scene_0016.png CHANGED

Git LFS Details

  • SHA256: ce3bd4dc35078a6b83ca5fb06bab36cfbab43b20590c12d094a80c312a0dd12e
  • Pointer size: 131 Bytes
  • Size of remote file: 710 kB

Git LFS Details

  • SHA256: 145628a885d11a198f38cb2ec933e9ef4676737ee2925bc18d93c36926a68051
  • Pointer size: 131 Bytes
  • Size of remote file: 503 kB
preview/0002_scene_0059.png CHANGED

Git LFS Details

  • SHA256: 6a40f24de03e270e24a212689b7b7136a87f7dba936f9c254ea69c3eb422666b
  • Pointer size: 131 Bytes
  • Size of remote file: 616 kB

Git LFS Details

  • SHA256: 3feaee6a45be2457590af723ee9f514a52bf92123ef04150909c4c4fa34298fa
  • Pointer size: 131 Bytes
  • Size of remote file: 632 kB
preview/0003_scene_0001.png CHANGED

Git LFS Details

  • SHA256: 7b3a1defa3f0ac93e903ad1e051477187a172e609c7ccce79b5cf0a09e1f93b8
  • Pointer size: 131 Bytes
  • Size of remote file: 629 kB

Git LFS Details

  • SHA256: 00b739ee10f281de1b46a8d6d284e31b6dc2c7bc2cff41d1873cbea0c56101a2
  • Pointer size: 131 Bytes
  • Size of remote file: 767 kB
preview/0004_scene_0004.png CHANGED

Git LFS Details

  • SHA256: 3cbba59e777300d314961a3e04140c86e0c3b90a2581183853e755f8cd44c7ae
  • Pointer size: 131 Bytes
  • Size of remote file: 847 kB

Git LFS Details

  • SHA256: c5e2706a287104d4cc9b80dffe446df1b901830e48c945ac00e179a5d3208b1a
  • Pointer size: 131 Bytes
  • Size of remote file: 781 kB