Source-linked AI summary
Catalogue Photography as a Cold Start: Toward Deployable Carbide Burr Recognition
Abilash Philip Madavath, Chandra Yuvesh Aubeeluck, Augustin Raju, Nicolas Pyschny, Felix Hackelöer, Florian Zwanzig
TL;DR
The paper asks whether catalogue photographs can support burr recognition when no labelled field imagery exists. It evaluates attribute-based representations and retrieval strategies under catalogue-to-field domain shift, finding that trained decomposition works strongly on catalogue images but transfers incompletely, while grayscale conversion and order-sheet constraints provide the largest gains. Catalogue photography is therefore a useful cold start rather than a deployment-ready training domain.
Problem
The paper addresses how to recognize manufactured carbide burrs when catalogue photography is the only available supervision and deployment imagery differs from it.
Method
The study compares catalogue-trained representations and scoring strategies for standardized head shape and tooth profile, evaluating them on held-out field photographs and known-order assignment.
Results
Training yields ARI 0.94–0.97 on catalogue images, while grayscale conversion adds +0.22 and order-sheet Hungarian matching adds +0.11 under field transfer.
Takeaways & Limitations
Catalogue photography provides a day-one baseline, but reducing domain sensitivity through grayscale inputs and order-sheet constraints matters more than increasing model complexity.
Takeaways & Limitations
The field set is small and incomplete, and the photographed tools are not packaging-line images, so the reported domain gap is a lower bound.
Abstract
from arXiv · showhide
Verifying that manufactured batches of milling tools or carbide rotary burrs conform to production order sheets remains a largely manual and error-prone quality assurance task. Automating this process with computer vision faces a critical cold-start constraint since no labelled imagery is available, leaving manufacturer catalogue photography as the sole source of supervision. We investigate how far catalogue supervision can support an industrial recognition pipeline under domain shift, explicitly measuring the gap between catalogue separability and performance on held-out field photographs. Our findings reveal three key insights. First, off-the-shelf frozen feature extractors do not reliably separate the two task attributes, head shape and tooth profile, motivating targeted representation learning. Second, metric learning produces near-perfect unsupervised cluster discovery on catalogue images (adjusted Rand index 0.94--0.97), but less than half of this gain transfers to field photographs. Third, the largest transfer gains do not come from model scale or representation complexity, but from simple changes that reduce domain sensitivity: converting images to grayscale (+0.22) and constraining retrieval using the known order sheet via Hungarian assignment (+0.11). We therefore treat catalogue photography as a useful cold start rather than a deployment-ready training domain, and provide empirical baselines and an evaluation protocol for catalogue-to-field transfer in precision tool manufacturing.
I. INTRODUCTION
The paper examines whether manufacturer catalogue photographs can cold-start burr recognition when field imagery is unavailable, while measuring the catalogue-to-field gap. It represents tools through standardized head-shape and tooth-profile attributes and explicitly limits the study to catalogue-only training and evaluation on held-out field photographs.
- Motivation: Manual inspection of carbide burr batches is error-prone because visually similar tools can differ in tooth pitch or cut direction.The task is to verify pallet contents against scanned production order sheets across hundreds of burr variants.
- Motivation: Catalogue photography is the only available training source before installation, but its studio rendering differs systematically from pallet-camera imagery.Specular reflections and line lighting dominate the field setting, creating a catalogue-to-field domain shift.
- Approach: The study describes each tool by independent head shape and tooth profile rather than thousands of combined types.These properties are standardized by DIN 8032 and DIN 8033, and the paper tests whether this decomposition transfers to real photographs.
- Protocol: The evaluation trains exclusively on catalogue images and reserves field photographs for held-out assessment, excluding camera integration and continual learning.This protocol treats catalogue-only learning as a cold start with no deployment-domain exposure during training, validation, or model selection.
II. TASK, DATA AND PROTOCOL
The pipeline recognises scale-normalised burr head crops using head shape and tooth profile, while accounting for label structure and cross-shape appearance changes.
- Burrs are represented by 9 DIN 8032 head shapes and 15 DIN 8033 tooth profiles.
- The pipeline predicts only the attribute pair (s, p), because diameter and shank length are better resolved through scale information or the order sheet.
- The label space is sparse and imbalanced, with existing combinations ranging from fewer than ten images to over a hundred.
- The same tooth profile can look different across head shapes, requiring profile-based models to learn shape-invariant representations.
B. The order sheet as a constraint
The known order sheet converts catalogue matching from open-set recognition into assignment to expected articles, but dataset construction and field coverage impose important constraints.
- The order sheet defines a known list of expected articles, allowing Hungarian assignment of observed tools using per-attribute similarity scores.This is evaluated as one scoring option rather than assumed to be universally optimal.
- The catalogue contains 770 shape-labelled images and 838 profile-labelled images, with 724 photographs carrying both labels.Images were scraped from manufacturer pages, cropped to the head, orientation-normalised, and filtered to classes with at least five images.
- The field data comprise 45 shape photographs and 52 profile photographs, covering only 7 of 9 shapes and 7 of 15 profiles.Because the pools have no jointly labelled images, combined performance is reported only as an independence-based product.
- Grouping images by parsed tool identity removed catalogue split leakage from repeated surface treatments, reducing inflated catalogue accuracy from about 0.97 toward valid post-fix results.The field set was near 0.51 before the split correction, illustrating the effect of near-duplicates across partitions.
E. Metrics
The evaluation combines local retrieval, global clustering, recognition accuracy, and catalogue-to-field degradation, while treating the small field set as a domain-shift diagnostic rather than a ranking benchmark.
- R@1 measures whether the nearest feature neighbour shares the label, whereas ARI measures global alignment between clusters and true classes.The two metrics can disagree because they assess local neighbourhoods versus the overall feature-space structure.
- Top-1 accuracy is the standard recognition score, while field gap Δacc = acccatalogue−accfield measures rendering-style sensitivity; lower gap is better.Ranking configurations by catalogue accuracy alone can reward rendering similarity instead of tool identity.
- With only 45 and 52 field images, single-accuracy intervals are roughly ±0.04–0.09 and the leading configurations are statistically equivalent by paired McNemar testing.The field set is therefore used to detect domain shift rather than rank the top approaches.
III. FOUR STAGES
Stage 1 tests whether frozen, off-the-shelf feature extractors can separate burr head shape and tooth profile without task-specific training. No single representation handles both attributes reliably.
- 41 feature extractors were evaluated with six clustering algorithms across the two burr attributes.The study covered self-supervised transformers, supervised ImageNet CNNs and ViTs, and a geometric contour descriptor.
- DINOv3 ViT-L/16 led shape separation with ARI 0.34 and R@1 0.89, while EfficientNet-B4 led profile separation with ARI 0.33.DINOv3’s profile ARI fell to 0.10, showing attribute-specific strengths.
- The shape and profile leaderboards shared no entry, so no frozen representation reliably separated both attributes.Scores across EfficientNet and ConvNeXt scales varied without a consistent trend, and HDBSCAN never won a row.
- The best frozen configuration became baseline A1 for subsequent comparisons.
B. Stage 2: what training does to the features
Training specialized representations transforms catalogue feature geometry: the two-stream model creates attribute-specific clusters that fixed pretrained features do not provide. Global clustering structure and local retrieval can nevertheless diverge.
- A4 trains two separate backbones with additive angular margin loss, one for shape and one for profile.The clustering protocol remains unchanged from Stage 1.
- A4 reaches ARI 0.94–0.97 across all four clustering algorithms, while HDBSCAN independently recovers 9 shape and 15 profile clusters.The UMAP projection shows tight, well-separated groups for each stream’s own attribute.
- A4’s shape and profile streams remain specialized, scoring ARI below 0.03 against the opposite attribute.This cross-check indicates that each stream encodes its intended attribute rather than both attributes indiscriminately.
- A2 reaches R@1 0.96–1.00 while its profile ARI remains 0.34, demonstrating that local retrieval can coexist with fragmented global structure.HDBSCAN splits the A2 profile space into 71–89 fragments instead of 15 groups.
C. Stage 3: what survives real photographs
Stage 3 measures how catalogue-trained representations transfer to held-out field photographs. Training improves field performance, but catalogue gains substantially overstate the improvement that survives domain shift.
- A4 attains the top nominal accuracy on both catalogue and field splits, but A3 performs closely behind.The comparison is based on means over three seeds and two colour modes; a larger field pool is needed to establish a statistically significant ranking.
- Nearly 0.50 catalogue improvement from training transfers as only about 0.20 on real photographs, with under half the gain surviving.The field gap ∆acc grows five-fold relative to the catalogue comparison.
- An edge-map encoder reaches catalogue profile ARI 0.85–0.91 yet performs below the untrained baseline in the field.This provides a sharp example of catalogue separability failing to transfer.
- A4’s field streams score 0.91 on shape and 0.64 on profile while scoring 0.22 on the opposite attribute in each case.This specialization supports approximately independent per-attribute costs for order-sheet matching.
D. Stage 4: what actually moved the field number
Eight no-retraining changes were tested against the incumbent on identical field images, but only grayscale conversion and order-sheet constraints produced meaningful gains. The best postprocessed frozen setup remained below A4, while the order-sheet result is limited to frozen features.
- Stage 4: Only two changes survived comparison against measurement noise: grayscale conversion and order-sheet-constrained retrieval.Resolution and pooling were indistinguishable from noise, and alternative backbones generally lost accuracy.
- Stage 4: The best postprocessed frozen configuration reached 0.48 combined, below A4’s 0.56 even on the easier constrained task.This indicates inference-time constraints alone did not replace representation learning.
- Stage 4: The order-sheet gain was measured only on frozen features, so its transfer to the trained two-stream system remains unresolved.Both surviving changes leave the representation unchanged.
IV. DISCUSSION
The main transfer improvement came from removing catalogue colour cues, while operational constraints offer additional headroom but do not yet support autonomous inspection. Several negative results reflect dataset or implementation limitations rather than definitive method failures.
- Discussion: +0.22 was the largest performance boost, obtained by converting images to grayscale.Catalogue colour grading acts as a shortcut that creates a chromatic shift unavailable on the factory line.
- Discussion: Order-sheet bipartite matching provided a +0.11 gain, but combined field accuracy near 0.56 remains below the over-99% reliability required to replace manual inspection.The paper frames the resulting mid-60% operational accuracy as headroom rather than autonomous deployment.
- Discussion: The cascade underperformed the parallel two-stream design because per-shape data was too thin, while CORAL and augmented-gallery expansion failed under their tested implementations.The authors explicitly avoid treating these outcomes as verdicts on the underlying methods.
- Discussion: Catalogue style-invariance testing clustered by tool shape because product series and photography setups were confounded with geometry.Style agreement tracked shape agreement almost one-to-one, while profile separation remained poor.
V. LIMITATIONS AND CONCLUSION
The field evaluation constrains confidence in both rankings and combined accuracy, while the broader conclusion is that catalogue images provide a useful cold start but not a deployment-ready domain. Domain shift, rather than model capacity alone, remains the central boundary.
- Limitations: At 45 and 52 field images, the evaluation cannot separate leading approaches or support honest deployment confidence intervals until missing classes are covered.The paper prioritizes expanding the field set before further model work.
- Limitations: A paired field set with both attributes labelled would test the independence assumption behind combined numbers and expose correlated failures.Current field photographs are not packaging-line images, so the reported gap is a lower bound.
- Conclusion: Catalogue photography is a usable cold start, but under half of training’s gain survives real photographs and the main transfer improvements remove colour or use the order sheet rather than changing the model.The conclusion identifies domain shift, not model capacity, as the central obstacle.