Benchmark Dataset · IT Professional, IEEE · 2026
A photometrically normalized, multiview face dataset that flattens 3-D geometry into six canonical 2.5-D projections — ready to train on, with no mesh handling.
CRBP reconstructs a textured 3-D mesh for each of the 105 Bosphorus subjects, renders six fixed camera views per scan, and normalizes each face to a single subject-specific mean colour — trading fine texture for photometric consistency so geometry carries the signal. Every projection is then passed through a fine-tuned YOLOv8 detector to produce a tightly cropped variant.
The six mean-colour projections rendered from each mesh, background retained. Better suited to tasks that need wide view coverage and real-world variability.
YOLOv8-localized crops that emphasize the eyes, nose, and mouth. Cleaner alignment for CNN training — at the cost of some pose- and occlusion-heavy samples.
How CRBP compares
Existing 3-D face datasets each miss something for image-based learning. CRBP is built to check every box.
| Dataset | 3-D geometry | Texture quality | Multiview | Colour consistency | Occlusion labels | Facial crops | 2.5-D projections |
|---|---|---|---|---|---|---|---|
| BU-3DFE | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ |
| Texas 3D | ✓ | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ |
| FRGC v2.0 | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ |
| Bosphorus | ✓ | ✗ | ✗ | ✗ | ✓ | ✗ | ✗ |
| 4DFAB | ✓ | ✓ | ✗ | ✓ | ✗ | ✗ | ✗ |
| FaceScape | ✓ | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ |
| NoW | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ |
| CRBP (ours) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
Table 1. Comparison of 3-D facial datasets relevant to 2.5-D analysis (from the paper).
One pass over the Bosphorus database turns each raw scan into two ready-to-train subsets. No mesh handling downstream — you consume PNGs.
Fig 1. The CRBP generation pipeline (Algorithm 1). Raw renders are kept as CRBP-Raw; YOLOv8-cropped renders become CRBP-YOLOv8.
CRBP's core move is a collapse: a 3-D mesh, rotated to a fixed viewpoint, rendered onto a 2-D plane. Pick a viewpoint to rotate the subject, drag to orbit, and watch where the camera ray meets the projection plane. Toggle Real mesh to swap the schematic for one de-identified Bosphorus scan, shown as geometry only. This viewer illustrates the projection method — it is not a figure from the paper or a dataset file.
Each viewpoint is a fixed (azimuth, elevation) pair. This is the table the published paper points to but that lives only in the supplementary file — reproduced here so the geometry is reproducible without the download.
| Viewpoint | Azimuth | Elevation |
|---|---|---|
| Frontal | 0° | 0° |
| Elevated frontal | 0° | 45° |
| Right three-quarter | 45° | 45° |
| Right profile | 90° | 0° |
| Left three-quarter | 135° | 45° |
| Left profile | 270° | 0° |
The point of shipping both subsets is that they behave differently. YOLOv8 cropping sharpens the easy cases and quietly discards the hard ones — which is exactly what makes the raw set worth keeping. Every figure below is drawn from the released code and results.
YOLOv8 favours frontal, unoccluded faces. Yaw-rotation samples fall from 4,410 to 1,826, and hair-/mouth-occluded faces are heavily under-represented after cropping.
Fig. 2 · Fig. A.3–A.4On a 105-way identity benchmark with EfficientNet-B0, the cropped set reaches 86.0% accuracy and 0.859 MCC, versus 82.9% and 0.827 on the raw set.
Fig. A.6–A.7Replacing each face's texture with one mean colour removes appearance shortcuts, so the projections are a cleaner test of whether a model reads 3-D shape.
Eq. 1 · Fig. 1Cleaner numbers on the cropped set come partly from dropping difficult views. Raw stays the honest choice for pose-robust and occlusion-aware evaluation.
Fig. A.5Sample count after each stage
Rendering multiplies every category by 6; YOLOv8 then drops the hardest views unevenly — pose and occlusion classes lose the most.
Fig 2. Per-category sample counts across the three stages (excludes invalid scans). Source: 01_CRBP_EDA_and_Visualization.ipynb.
Identity accuracy · Raw vs YOLOv8
EfficientNet-B0, 105 subjects, 60/20/20 split, 10 epochs.
Fig 3. Test-set metrics. Source: 02_..._EfficientNetb0.ipynb.
Training loss over epochs
Both converge; the cropped set trains lower and faster.
Fig 4. Per-epoch training loss from the live run.
Full results · EfficientNet-B0
105-way identity recognition, 60/20/20 split, 10 epochs. Best value per row in bronze.
| Metric | CRBP-Raw | CRBP-YOLOv8 |
|---|---|---|
| Accuracy | 0.829 | 0.860 |
| Balanced accuracy | 0.814 | 0.847 |
| Precision | 0.837 | 0.867 |
| Recall | 0.829 | 0.860 |
| F1 score | 0.829 | 0.861 |
| MCC | 0.827 | 0.859 |
| AUC ROC | 0.996 | 0.997 |
Table 2. Test-set metrics for both subsets. YOLOv8 cropping wins across the board — partly because it drops the hardest views. Source: 02_..._EfficientNetb0.ipynb.
Because CRBP normalizes appearance and standardizes viewpoints, it suits tasks that depend on facial geometry rather than photographic texture.
Expression and action-unit categories are preserved across all six views.
105-subject identity labels with pose and occlusion variation built in.
Consistent multiview projections for building or texturing 3-D face models.
Facial-geometry cues studied in anxiety, dementia, and Parkinson's research.
Attention and affect estimation from standardized, appearance-normalized faces.
Labelled hard cases — yaw, pitch, glasses, hair — for stress-testing models.
The projections are distributed through IEEE DataPort. Because CRBP is derived from the Bosphorus 3D Face Database, its use is subject to Bosphorus's own licence — please secure access to Bosphorus and cite it alongside CRBP.
@article{sharma2026crbp,
author = {Sharma, Sahil and Sood, Ashima and Kumar, Vijay},
title = {{CRBP}: A Benchmark Dataset of Color-Rendered
3-D-to-2.5-D Facial Projections},
journal = {IT Professional},
year = {2026},
volume = {28},
number = {1},
pages = {59--65},
doi = {10.1109/MITP.2025.3624674}
}
@data{sharma_crbp_dataset,
author = {Sharma, Sahil and Sood, Ashima and Kumar, Vijay},
title = {Colour-Rendered Bosphorus Projections Database},
publisher = {IEEE Dataport},
year = {2024},
doi = {10.21227/4qns-ys45}
}