Deep Reading Notes: Shape-Biased Texture Agnostic Representations for Improved Textureless and Metallic Object Detection and 6D Pose Estimation

1. Basic information

• Short name: Randomized Texturing / Shape Bias
• Authors: Peter Hönig, Stefan Thalhammer, Jean-Baptiste Weibel, Matthias Hirschmanner, Markus Vincze
• Affiliations: TU Wien, UAS Technikum Vienna
• Venue: WACV 2025
• Published paper: CVF Open Access
• Code: randomized_texturing
• Reading mode: Deep Dive

2. Conclusions first

The paper frames textureless / metallic synthetic-to-real transfer as a training shortcut: fixed textures or low-frequency solid colors let CNNs exploit appearance correlations instead of contours, surface projections, occlusion, and shadows. Without architecture changes, BlenderProc PBR replaces uniform colors with randomized natural-texture UV mapping. The 1,226 inconsistent appearances disrupt this shortcut.

Experiments give bounded support, rather than a universal proof that textures are irrelevant:

1. Fully shape-biased detection / pose raises T-LESS GDR-Net BOP AR from 56.29 to 58.32, and Pix2Pose from 46.68 to 50.13;
2. ITODD improves from 10.70 to 11.10 and 12.90 to 15.30 respectively, with low absolute accuracy;
3. Gaussian / shot noise and brightness improve; motion / Gaussian blur barely do. This agrees with reducing appearance shortcuts without recovering lost geometric detail;
4. Detector / pose interactions matter: T-LESS Pix2Pose barely changes with pose-only bias, needing shape-biased YOLOx too for its total 3.45 AR gain;
5. Accurate CAD matters more than texture intervention: Pix2Pose CAD + shape bias reaches 62.95 AR, versus reconstructed mesh + shape bias 46.57;
6. Some older detectors and augmentation configurations decline, ruling out an untuned universal augmentation;
7. Evidence covers CNNs, per-object instance-level training, and T-LESS / ITODD, without establishing ViT, diffusion-feature, category-level, or unseen-object gains.

The reusable experimental lesson is to intervene on material distributions while controlling detectors, pose estimators, image corruption, online augmentation, and geometry priors separately.

2.1 Key figures and levels of evidence

Evidence	Controlled variables	Supported conclusion	Unsupported conclusion
Figure 2	Only object color / texture generation changes	Easy integration into standard BOP PBR	Networks necessarily learn explicit 3D shapes
Table 3	T-LESS classification, same ResNet50	UV mapping exceeds the reproduced AdaIN setting	Superiority over all style-transfer methods
Table 4	Two datasets / pose models, YOLOx detections	Matching or improved tested 6D pose results	Universal gains across detection / pose architectures
Figure 6	T-LESS, GT detections, GDR-Net	Better noise / brightness robustness, little blur benefit	Equal gains under detection errors or every real corruption
Figure 7	Pix2Pose / YOLOx, T-LESS	Augmentation / shape-bias hyperparameter interactions	One universally optimal augmentation strength
Table 6	Pix2Pose, T-LESS, GT detections	Appearance randomization cannot replace accurate geometry	Reconstructed meshes are never useful

3. Task definitions and terminology

3.1 What pose is estimated?

The paper represents instance-level 6D object pose as:

$$\mathbf{T}_{CO} = [\mathbf{R}\mid\mathbf{t}],$$

Here $\mathbf{R}\in SO(3)$ and $\mathbf{t}\in\mathbb{R}^3$ map object coordinates to camera coordinates. Training / testing use T-LESS / ITODD instances, with a separate GDR-Net or Pix2Pose estimator trained per object.

It therefore differs from:

• category-level pose;
• model-free / reference-free pose;
• unseen-object generalization;
• Orientation-only or relative-rotation tasks.

BOP VSD, MSSD, and MSPD Average Recall evaluate full poses; CAD / reconstructed meshes remain necessary to generate training data and define geometry.

3.2 What is texture bias?

Texture bias is CNN preference for local texture statistics among available cues. ImageNet classification may succeed through such cues, but textureless industrial pose can suffer brittle shortcuts:

• Continuous, stable synthetic object colors;
• Real appearance altered by light, exposure, specularity, and shadows;
• Learned local appearances inconsistent with pose-discriminative geometry;
• Failed synthetic-to-real transfer.

3.3 What does shape bias mean here?

Training distributions favor shape-related evidence: silhouettes, boundaries, occlusion structure, surface texture deformation, and PBR lighting / shadows. There is no explicit 3D-coordinate supervision or proof of a complete decodable shape in feature maps.

A stricter interpretation is:

Randomized texturing reduces statistical coupling between fixed appearance and labels / poses, making cross-texture geometric projection cues more useful for minimizing risk.

Texture agnosticism similarly means relative robustness, without mathematical texture invariance.

4. A training-distribution interpretation

Let mesh be $M$, category / instance $o$, pose $p$, camera / lighting / background nuisance variables $z$, and surface appearance $a$. Rendering produces:

$$I = \mathcal{R}(M_o, p, z, a).$$

The texture-biased baseline uniformly samples RGB colors for untextured objects:

$$a = (r,g,b), \qquad r,g,b\sim\mathcal{U}(0.1,0.9).$$

Changing colors still leaves large low-frequency continuous materials within each object. Networks can learn surface statistics specific to the synthetic pipeline.

The proposed texture pool is:

$$\mathcal{T}=\{T_1,T_2,\ldots,T_n\}, \qquad n=1226,$$

Each rendering samples uniformly:

$$T_i\sim\mathcal{U}(\mathcal{T}), \qquad a=\operatorname{UVMap}(M_o,T_i).$$

The training objective remains:

$$\min_\theta\;\mathbb{E}_{o,p,z,T_i} \left[\mathcal{L}\big(f_\theta(\mathcal{R}(M_o,p,z,T_i)),y(o,p)\big)\right].$$

When the same $o,p$ admits many $T_i$, fixed-texture features cannot consistently lower expected losses; contours, occlusion, and surface deformations that remain stable across $T_i$ become more valuable.

This resembles appearance intervention, but is not a strict causal experiment. Random textures jointly alter spectra, edge density, color distributions, and PBR interactions without isolating them.

5. Rendering and implementation

5.1 Relationship to BOP PBR

BlenderProc and standard BOP configurations retain randomization of:

• Background textures;
• Camera viewpoints;
• Object positions and rotations;
• Scene object combinations;
• Lights and shadows;
• Specularity and roughness.

The added variable is foreground surface texture. Background randomization helps foreground separation but leaves internal object appearance shortcuts intact.

5.2 Two material configurations

Parameter	Texture-biased baseline	Shape-biased variant
specularity	$\mathcal{U}(0,1)$	$\mathcal{U}(0,1)$
roughness	$\mathcal{U}(0,1)$	$\mathcal{U}(0,1)$
Surface color	Per-channel RGB $\mathcal{U}(0.1,0.9)$	Uniformly sample and UV-map 1,226 textures
Scene / camera / lighting	Standard BOP	Same

Both randomize specularity and roughness. The difference is solid-color versus spatial-texture priors, rather than extra reflection-parameter randomization.

5.3 Texture library

• The Creative Commons texture database used by the paper;
• Translucent textures removed;
• 1,226 opaque textures retained;
• A texture can cover multiple objects, and each object changes textures across renderings;
• Figure 2 shows only 5 examples, not the full training pool.

Low-cost UV mapping needs no manually designed semantic texture per industrial part. Its purpose is diverse surface patterns without stable identity associations.

5.4 Data scale and cost

The full synthetic set has 50,000 renderings. Sampling and mapping add negligible cost relative to standard PBR generation / training, with no new inference module or model FLOPs. Actual throughput still depends on texture I/O, material caching, and renderer settings, which reproduction should record.

6. Why might UV mapping suit pose better than style transfer?

AdaIN-style transfer redistributes 2D style statistics and can encourage classification shape bias; pose needs precise spatial cues. The authors offer three motivations:

1. Textures attach to actual 3D surfaces: rendering consistently handles perspective, surface deformation, occlusion, and shadows;
2. Textures decouple from identity: different appearances share pose labels;
3. One generation method supports classification, detection, segmentation, and pose, avoiding separate style-transfer pipelines per task.

Controlled T-LESS classification uses ResNet50 with GT box crops:

Training	Top-1 accuracy	Top-5 accuracy
Solid-color texture bias	63.22	86.27
AdaIN style-transfer shape bias	75.88	91.16
UV-mapping shape bias	87.27	97.37

Against AdaIN, UV mapping improves top-1 / top-5 relatively by 15.01% / 6.81%. This supports classification representations in this setting; pose evidence comes separately from Table 4.

7. Experimental design

7.1 Dataset selection

T-LESS

T-LESS has textureless industrial objects, geometric similarities, symmetries, occlusion, and clutter, testing whether fixed color priors obscure shape differences.

ITODD

ITODD’s metallic surfaces depend strongly on incident light, specularity, and exposure, directly stressing appearance changes.

ITODD test GT is not public. BOP-server results and Figure 4 without detection GT establish benchmark gains but cannot support comprehensive per-object / failure-type attribution. The conclusion also calls for fuller metallic-object evaluation.

7.2 Detectors

• Faster R-CNN;
• RetinaNet;
• YOLOx;
• YOLOv8;
• YOLOv10.

Models span two-stage / one-stage designs and generations, without transformer detectors. Newer YOLO models benefit more consistently; the authors suggest capacity and architecture, without causally controlling those factors.

7.3 Pose estimators

Model	Backbone	Protocol
GDR-Net	ResNet34	Separate estimator per object; YOLOx detections
Pix2Pose	ResNet50	Separate estimator per object; YOLOx detections

Both are CNN-based multistage instance-level systems. Boxes affect crops and subsequent correspondence / regression, so Table 4 separates object-detection bias (O.D.) from pose-estimation bias (P.E.).

7.4 Metrics

Overall BOP AR averages three terms:

$$AR_{BOP}=\frac{AR_{MSPD}+AR_{MSSD}+AR_{VSD}}{3}.$$

• MSPD: maximum symmetry-aware projection error, emphasizing image alignment;
• MSSD: maximum symmetry-aware surface distance, emphasizing 3D displacement;
• VSD: visible surface discrepancy, accounting for occlusion / visible regions;
• Each first averages recall across error thresholds.

ADD or angular errors alone may penalize equivalent symmetric poses unfairly; BOP’s three terms better suit textureless / symmetric objects.

8. Detection: gains are not universally monotonic

Detection findings are more nuanced than gains for all five detectors.

8.1 Default augmentation

Reported relative changes include:

• T-LESS: Faster R-CNN 247.03%, RetinaNet 200.87%. Low baselines explain these large relative gains; absolute mAP is essential context;
• ITODD: Faster R-CNN 11.37%, RetinaNet down 7.91%;
• T-LESS YOLOx / YOLOv8 / YOLOv10: 51.52%, 9.73%, 6.78%;
• ITODD’s three YOLO models: 60.31%, 44.91%, 28.49%.

8.2 Tuned online augmentation

After retuning online augmentation per setting:

• T-LESS Faster R-CNN / RetinaNet decline by 3.34% / 5.80%;
• ITODD counterparts improve by 0.87% / 6.51%;
• ITODD YOLOx / YOLOv8 / YOLOv10 improve by 6.03%, 46.20%, 27.47%.

Randomized texturing and augmentation can overlap or conflict. Strong existing color / contrast / noise augmentation can shrink marginal gains, while unsuitable strengths damage detection training.

The supported claim is more consistent benefits for newer YOLO models in tested settings, with no unconditional gain across all detectors.

9. Main 6D pose results

9.1 Complete numerical results

Dataset	Model	O.D.	P.E.	MSPD	MSSD	VSD	AR
T-LESS	GDR-Net	texture	texture	72.66	50.52	45.70	56.29
T-LESS	GDR-Net	texture	shape	73.54	52.94	48.09	58.19
T-LESS	GDR-Net	shape	shape	73.73	53.05	48.17	58.32
T-LESS	Pix2Pose	texture	texture	68.35	37.85	33.84	46.68
T-LESS	Pix2Pose	texture	shape	67.71	38.25	34.10	46.69
T-LESS	Pix2Pose	shape	shape	71.22	41.86	37.30	50.13
ITODD	GDR-Net	texture	texture	14.30	10.10	7.80	10.70
ITODD	GDR-Net	texture	shape	14.30	10.10	7.80	10.70
ITODD	GDR-Net	shape	shape	15.30	10.20	7.90	11.10
ITODD	Pix2Pose	texture	texture	22.50	8.00	8.00	12.90
ITODD	Pix2Pose	texture	shape	22.70	10.10	10.30	14.30
ITODD	Pix2Pose	shape	shape	23.10	11.00	11.70	15.30

9.2 Separate pose and detector contributions

T-LESS / GDR-Net

• baseline: 56.29;
• Pose-only change: 58.19, +1.90 AR, relative +3.38%;
• Also change detection: 58.32, another +0.13 AR;
• Full pipeline relative to baseline: +3.61%.

Most GDR-Net gains come from pose training, with a small extra YOLOx benefit.

T-LESS / Pix2Pose

• baseline: 46.68;
• Pose-only change: 46.69, nearly unchanged;
• Also change detection: 50.13, another +3.44 AR;
• Full pipeline relative to baseline: +7.39%.

Pix2Pose shows why final gains cannot all be assigned to shape-biased pose representations: improved / distribution-consistent detections are a crucial mediator here.

ITODD / GDR-Net

• baseline: 10.70;
• Pose-only change: 10.70, unchanged;
• Also change detection: 11.10, +0.40 AR;
• Full pipeline relative to baseline: +3.74%.

ITODD / Pix2Pose

• baseline: 12.90;
• Pose-only change: 14.30, +1.40 AR, relative +10.85%;
• Also change detection: 15.30, another +1.00 AR;
• Full pipeline relative to baseline: +18.60%.

The relative gain is striking, but final AR is only 15.30. Also report the absolute +2.40 AR, so 18.60% does not suggest a solved problem.

9.3 What do component metrics show?

• T-LESS / GDR-Net: MSPD, MSSD, and VSD improve, extending beyond box alignment;
• T-LESS / Pix2Pose: MSPD 68.35→71.22, MSSD 37.85→41.86, VSD 33.84→37.30 all improve;
• ITODD / GDR-Net: gains mainly come from MSPD 14.30→15.30, with MSSD / VSD only +0.10;
• ITODD / Pix2Pose: larger relative MSSD / VSD gains suggest surface alignment improvements beyond detector crops.

Table 4 lacks uncertainty, multiseed variance, and per-object breakdowns, leaving unclear whether many small gains or a few difficult objects drive the means.

10. Image perturbations: the clearest mechanism boundaries

10.1 Protocol

• dataset: T-LESS;
• estimator: GDR-Net;
• One pose model per object;
• Ground-truth detections isolate detector error;
• Test-image corruptions vary in severity;
• Compare texture-biased and shape-biased training.

10.2 Gaussian and shot noise

Baseline AR collapses at low severity; the shape-biased model declines more smoothly. Richer high-frequency training patterns may reduce dependence on specific stable pixel statistics.

This mechanism remains tentative: no frequency-response, feature-attribution, or linear-probe experiment directly tests it.

10.3 Brightness

Shape-biased AR stays substantially higher at medium / high brightness severity. This matches the metallic-object motivation: features less tied to fixed colors / intensities should resist global brightness shifts.

10.4 Motion and Gaussian blur

Both models follow similar declines. Blur removes contours, edges, and fine spatial structure that shape-biased models also need. Changing training textures cannot recover geometric information absent from inputs.

This useful negative result prevents claims of universal corruption robustness.

10.5 Protocol limits

GT detections isolate pose backbones under idealized conditions. Real noise / brightness changes also damage detection, so Figure 6 does not establish equal full-pipeline advantages at those severities.

11. Ablation studies

11.1 Texture count

Figure 5 measures YOLOx detection mAP versus texture-pool size on T-LESS:

• $n=1226$ performs best;
• $n=5$ already reaches 93.85% of the maximum;
• $n=100$ reaches 99.63%.

Diminishing returns suggest few textures already disrupt fixed-color identity shortcuts; further textures broaden local patterns and spectral coverage.

These ratios belong to the YOLOx / T-LESS detector ablation. They do not establish that all pose models reach 99.63% with 100 textures.

11.2 Shape bias and online augmentation

The authors multiply ordinary augmentation probabilities by scale $\lambda\in[0,1]$ and grid-search it. Candidates include coarse dropout, Gaussian blur, sharpness, contrast, brightness, color, add, invert, multiply, Gaussian noise, and linear contrast. Detection / pose lists and base probabilities differ.

Figure 7 finds:

• Pix2Pose pose peaks around $0.2\leq\lambda\leq0.4$;
• YOLOx detection peaks around $\lambda=0.4$;
• Shape bias benefits pose across multiple strengths;
• Detection needs strength retuning for best mAP.

The two randomizations can be:

1. Complementary: surfaces change materials, online augmentation changes imaging;
2. Redundant: both alter color, contrast, and local statistics;
3. Conflicting: excessive augmentation destroys texture and shape cues together.

11.3 Mesh origin: appearance randomization cannot remove geometric errors

Geometry prior	Appearance prior	AR
reconstructed mesh	reconstructed texture	42.64
reconstructed mesh	randomized texture / shape	46.57
CAD	randomized grayscale color	59.24
CAD	randomized texture / shape	62.95

Two orthogonal effects are evident:

• Within each mesh type, shape bias adds +3.93 and +3.71 AR;
• Within each bias type, CAD exceeds reconstruction by +16.60 and +16.38 AR.

Geometry-prior quality is the larger main effect, with randomized textures adding gains. Errors in reconstruction scale, contours, or local structure cause systematic coordinate-supervision bias that texture randomization cannot repair.

12. Why do models benefit differently?

The authors attribute larger Pix2Pose gains to its higher-capacity ResNet50 versus GDR-Net’s ResNet34. Plausible, but experiments also mix:

• Heads and training losses;
• Correspondence / regression mechanisms;
• Crop and detector sensitivities;
• Baseline absolute AR;
• Online augmentation protocols.

Two models cannot establish that larger backbones necessarily gain more. Isolate backbone scale within one architecture or run capacity-matched comparisons.

A stronger observation is that gains jointly depend on local appearance use, crop sensitivity, and consistency between training distributions.

13. Main strengths

13.1 Modular intervention

No inference branch, changed pose output, or extra real labels are required. Integration needs a synthetic renderer capable of assigning mesh materials.

13.2 Matches textureless / metallic-object motivation

Real appearance depends on light, making fixed synthetic-color priors fragile. Unrealistic random textures deliberately broaden appearance support to encourage stable cues.

13.3 Evidence beyond clean accuracy

Corruptions, texture counts, augmentation strength, and mesh origin test where / why the method works, providing stronger evidence than one result table.

13.4 Separate detector and pose controls

Table 4’s bias combinations expose mediators in two-stage systems. T-LESS Pix2Pose shows why comparing only shape/shape with texture/texture would incorrectly credit pose estimation for much of the detector gain.

14. Limitations and causal boundaries

14.1 CNNs only

Vision Transformers and diffusion models are explicitly untested. DINO / ViT may prefer different shape / texture cues; gains, texture counts, and augmentation strengths need fresh evaluation.

14.2 Known instances with per-object training

Per-object networks can encode geometry in weights. Unseen-object systems use cross-instance matching or CAD-template retrieval, where random textures may help or harm reference-query matching.

14.3 Incomplete ITODD evidence

Without public test GT, the paper cannot perform comprehensive error analysis. Metallic objects motivate the title, but most diagnostic robustness / mesh ablations use T-LESS.

14.4 Missing statistical uncertainty

Main tables lack multiseed means, standard deviations, or confidence intervals. Full-pipeline T-LESS GDR-Net gains only 0.13 AR over pose-only bias; whether this exceeds training randomness remains unknown.

14.5 Shape representations are not measured directly

Accuracy and robustness show behavior, rather than probing feature geometry. Further tests could include:

• texture-swap consistency;
• silhouette / edge occlusion sensitivity;
• feature linear probe for object coordinates / normals;
• cross-texture correspondence stability;
• Centered kernel alignment between representations trained on different distributions.

14.6 Random textures may hurt fine-texture discrimination

T-LESS / ITODD lack stable textures. For similarly shaped products distinguished by logos / prints, enforced texture agnosticism can remove useful evidence. Mixing shape-biased and texture-faithful data may suit these better than total replacement.

14.7 Geometry, detection, and blur remain

• Reconstructed geometry still incurs about a 16 AR gap;
• Detector augmentation needs retuning;
• Motion / Gaussian blur does not improve;
• ITODD absolute AR remains low.

This addresses one data-distribution dimension within an industrial pose system.

15. Reproduction checklist

15.1 Data preparation

1. Obtain T-LESS / ITODD CAD and BOP camera / pose annotations;
2. Prepare a license-compatible texture library;
3. Remove alpha / translucent textures and textures the renderer cannot handle reliably;
4. Fix pool versions / file lists and record $n$;
5. Check each object’s UV mapping for extreme scales or unmapped regions.

15.2 Strictly comparable rendering variants

Texture baseline:

• Uniform RGB channel sampling from 0.1–0.9;
• Specularity / roughness sampled from 0–1.

Shape variant:

• Identical specularity / roughness;
• Uniform sampling from the texture pool;
• UV mapping onto object surfaces.

Match scene composition, cameras, lights, backgrounds, object / rendering counts, resolution, and seed policy; otherwise extra randomization differences confound results.

15.3 Training

• Match initialization policies, epochs, optimizers, and batch schedules;
• Tune reasonable online augmentation separately for both variants;
• Train detection and pose separately; retain at least texture/texture, texture/shape, and shape/shape combinations;
• Add shape detection + texture pose if affordable, completing the interaction analysis;
• Run at least 3 seeds and report means / standard deviations.

15.4 Evaluation

• clean BOP AR, AR_MSPD, AR_MSSD, AR_VSD;
• Detection mAP / recall, preventing detector effects from hiding pose contributions;
• Pose-only results with GT detections;
• End-to-end results with predicted detections;
• noise, brightness, motion blur, Gaussian blur;
• Per-object and symmetry-group breakdowns;
• runtime / rendering overhead;
• Texture-count / augmentation-strength ablations;
• Separate CAD and reconstruction results.

15.5 Common reproduction mistakes

• Randomizing backgrounds without foreground surfaces;
• Omitting baseline specularity / roughness randomization;
• Mistaking 5 example textures for the full pool;
• Reusing baseline-optimal augmentation without retuning shape variants;
• Reporting relative percentages without absolute AR;
• Extrapolating GT-detection robustness to full pipelines;
• Treating ITODD qualitative box overlays as quantitative evidence with public detection GT;
• Describing texture agnosticism as complete invariance.

16. Future research takeaways

These extensions follow from the evidence; the authors have not already validated them.

16.1 Adaptive texture–shape mixtures

Increase randomized textures for textureless / metallic objects, preserving faithful appearances for informative logos / printing. Mixture probabilities could depend on material and geometric symmetry:

$$I\sim \begin{cases} \mathcal{R}(M,T_{random}), & q(o)\\ \mathcal{R}(M,T_{faithful}), & 1-q(o). \end{cases}$$

This discourages appearance shortcuts while retaining discriminative textures.

16.2 Combine with physical material randomization

Roughness / specularity already vary, but textures are ordinary opaque surfaces. Further cover metallic BRDFs, anisotropy, clear coats, environment maps, and exposure responses; ablate random patterns separately from physical materials.

16.3 Revalidate foundation features

Repeat texture-swap consistency and BOP tests for DINOv2, DINOv3, CLIP / SigLIP, diffusion correspondences, and ViT detectors. Check existing shape bias and remaining texture-randomization gains.

16.4 Feature consistency beyond data intervention

Render multiple textures for the same object / pose, adding feature consistency to task loss:

$$\mathcal{L}_{inv}=\left\|f_\theta(I_{T_i})-f_\theta(I_{T_j})\right\|,$$

Align corresponding surface regions only. This directly encourages invariance but risks erasing useful real textures, requiring balance with a faithful branch.

16.5 Joint geometry-quality curriculum

Table 6 identifies geometry gaps as dominant. Randomize mesh noise, missing surfaces, normals, and textures to improve reconstruction-asset robustness, while preventing inaccurate geometric supervision from corrupting pose labels.

17. Relationship to 6D pose approaches

This is a training-data intervention upstream of correspondence, PnP, render-and-compare, or direct regression:

CAD / reconstructed mesh
        ↓
Material / texture distribution design ← this paper's contribution
        ↓
BlenderProc PBR synthetic images
        ↓
object detector + instance-level pose estimator
        ↓
BOP 6D pose metrics

Complementary relationships:

• Synthetic data: extends randomization to foreground surfaces;
• Correspondence: may encourage geometric 2D features, without direct correspondence evaluation;
• Model-based pose: still requires CAD / meshes;
• CAD-free reconstruction: improves reconstruction-based training without closing geometry gaps;
• Robustness: helps noise / brightness, with blur unresolved;
• Unseen-object pose: related motivation, but transfer remains untested.

18. Deep-reading conclusion

Randomized Texturing is a small intervention supported by carefully bounded experiments. The strongest conclusion is:

In tested CNN-based instance-level T-LESS / ITODD detection and pose pipelines, replacing uniform surface colors with many random UV textures reduces fixed-appearance dependence. Clean BOP AR matches or improves, and Gaussian / shot noise and brightness robustness improve substantially. Gains depend on detection, pose architecture, online augmentation, and geometry priors.

As an engineering option, it is inexpensive to try in BlenderProc / BOP, especially for textureless reflective parts. Research should add multiseed statistics, feature probes, ViT / diffusion tests, faithful-material mixtures, and full-pipeline corruption evaluation.

Three numbers keep expectations grounded:

1. Full T-LESS gains: GDR-Net +2.03 AR, Pix2Pose +3.45 AR;
2. ITODD Pix2Pose improves relatively by 18.60% but ends at only 15.30 AR;
3. CAD + shape bias exceeds reconstructed mesh + shape bias by 16.38 AR.

It addresses part of the appearance-shortcut problem within the broader challenges of industrial 6D pose.