acceptodds
Under review as a conference paper at ICLR 2027

CALIPER: Direct Point-Text Alignment for Dimension-Bearing CAD Retrieval

Abstract

Retrieving CAD models from a geometry-only gallery with natural-language text differs from open-vocabulary 3D retrieval in what the query contains. Engineers describe a part by its dimensions, so retrieval must match stated numbers against geometry. The prevailing design, which anchors a point encoder to a frozen CLIP text tower, discards this information on both sides, truncating expert captions before their dimensions and normalizing away the scale those dimensions refer to, and it freezes the tower that would have to learn them. This paper presents , a point–text bi-encoder for geometry-only CAD galleries that keeps the whole caption and each shape's corpus-frame scale and adapts both towers to paired CAD data. Ablations show that each preserved quantity and each adapted tower is load-bearing. Because the standard Text2CAD split leaks a majority of its evaluation shapes into training through near-duplicates, results are reported on a component-disjoint split, where reaches 10 0.955 under cluster credit and 0.848 by exact uid. Counterfactual probes establish aggregate-level grounding of the numbers in a query, and on human-written CADPrompt queries over held-out geometry retrieves dimension-bearing prompts seven times as often as the strongest zero-shot baseline.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multimodal representation learning, 3D shape retrieval, text-to-3D, contrastive learning, CAD, long-context text encoder, cross-modal alignment, benchmark contamination

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_z3ZPOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents CALIPER, a point-cloud-text dual encoder for CAD model retrieval with dimension annotations. This system adapts the pre-trained text encoder and point-cloud encoder through contrastive learning while retaining detailed CAD description information and geometric scale features at the level of the corpus. Additionally, this paper analyzes the "near duplicate data leakage" problem existing in the official Text2CAD data division and introduces a set of evaluation schemes based on component non-overlap.

Critical strengths
  1. This paper points out that there are differences between the general text-to-3D retrieval and CAD retrieval; in CAD retrieval, numerical dimensions are crucial for distinguishing parts with similar geometric shapes.

  2. The author discovered that there is a significant problem of near-duplicate data leakage in the official Text2CAD dataset. The component-disjoint split method proposed in the paper, as well as the introduction of two evaluation indicators, cluster-credit and exact-uid, have enhanced the rigor of the evaluation.

Critical weaknesses
  1. The Introduction and Method sections are presented in a largely pipeline-like, sequential manner, rather than providing a coherent and integrated description, analysis, and comparison of the overall problem. As a result, reviewers and readers must invest considerable effort in reconstructing the problem formulation and understanding the overall narrative of the paper. This fragmented presentation makes the manuscript unnecessarily difficult to follow and significantly undermines its readability.

  2. The proposed approach combines several established components, including pretrained text and point-cloud encoders, contrastive alignment, scale-preserving preprocessing, and staged fine-tuning. However, the method does not introduce a clearly novel architecture or learning objective. The primary contribution appears to lie in the empirical analysis and evaluation protocol rather than in the model itself. The authors should more clearly articulate the methodological novelty of their approach and explain how it advances beyond existing techniques.

3.The headline comparisons against the zero-shot Uni3D and OpenShape baselines involve substantial differences in domain adaptation and input coverage, which complicates the interpretation of the reported performance gains. Notably, the fully adapted Long-CLIP baseline with a 512-token input achieves an R@10 of 0.909, compared with 0.957 for CALIPER under the strict validation setting. This relatively small performance gap raises an important question: could comparable performance also be achieved by replacing the language encoder and applying supervised fine-tuning to the aforementioned zero-shot methods? The authors do not provide a sufficiently detailed analysis to address this possibility. Additional controlled experiments or a more thorough discussion are needed to establish whether the observed improvements stem from the proposed method itself or primarily from domain adaptation and fine-tuning.

4.The evidence supporting fine-grained dimensional grounding remains insufficient. Although the model exhibits strong sensitivity to uniform numerical perturbations, doubling an individual decimal value has little effect on retrieval performance, and swapping the length and width values does not lead to a measurable degradation. These findings raise concerns about whether the model genuinely grounds individual dimensional values and their semantic roles, rather than merely exploiting coarse-grained numerical patterns. The authors should provide further evidence to substantiate their claims regarding fine-grained dimensional understanding.

  1. The authors consulted Fusion360 geometry during model development, although they subsequently evaluated the model on sealed query sets. While this evaluation provides some evidence of transfer to held-out queries, it does not fully resolve concerns regarding broader generalization. In particular, the extent to which exposure to Fusion360 geometry during development may have influenced model design, data selection, or evaluation outcomes remains unclear. The authors should clarify the relationship between the development data and the evaluation sets and provide additional evidence of generalization to unseen geometries or domains.
Questions for authors

see above

1 comment
AuthorsOpenReview ↗

Reply to Reviewer z3ZP

Thank you for the review. We answer W1–W5 in order. Unless marked, retrieval numbers are cluster-credit R@10 on our strict validation split.

W1 — Presentation. We agree. The introduction and method follow the pipeline and leave the reader to reconstruct the problem and the argument. The paper needs to be restructured around the problem, what it requires, and how each requirement is tested.

W2 — Novelty. We agree that the method introduces no new architecture or objective, and that the contribution lies in the empirical analysis and the evaluation protocol. A new full-chain control also shows that the results do not depend on the cluster-aware loss (0.957 vs 0.958 with plain InfoNCE).

W3 — Could replacing the language encoder and fine-tuning reach comparable performance? Yes. We swapped the text tower and trained each variant with our full recipe. The table gives R@10; its first row is the untrained zero-shot baselines.

Model / text towerstrict teststrict validationCADPrompt, prompts with dimensions (n = 182)
zero-shot Uni3D-g / OpenShape-G140.004 / 0.004—0.027 / 0.016
Long-CLIP, 512 tokens0.9060.9090.291
gte on the identical 512-token text0.8940.8970.346
Long-CLIP stretched to 2,048 tokens0.9290.9410.291
TULIP, 2,048 tokens—0.954—
gte, 2,048 tokens (CALIPER)0.9550.9570.341

A dash marks a setting that was not evaluated. The CADPrompt column uses the seen-geometry protocol of our appendix table for the trained rows (the CADPrompt shapes are in their training data) and Table 6 for the zero-shot rows. TULIP is single-seed.

The adapted towers that read (almost) the whole caption land within about three points of CALIPER on machine captions and about five on human prompts. The improvement over the zero-shot baselines therefore comes from domain adaptation and input coverage, as you suspected, and not from the specific encoders. The zero-shot rows of Table 1 show how far off-the-shelf models are from this task; they are not a controlled comparison, and the paper should say so more clearly. We did not fine-tune Uni3D or OpenShape end to end. The closest evidence on the point side is that frozen Uni3D-g features with our heads outperform frozen Utonia (0.517 vs 0.471).

W4 — Fine-grained dimensional grounding. We agree that the evidence in the submission was insufficient, and one of its probes was flawed. New results:

  • Minimal pairs. On pairs of similar shapes that differ by ≥10% in one extent, forced-choice accuracy is 0.910, against 0.505 with the numbers stripped.
  • Role binding. The length↔width swap was confounded, since the swapped caption describes a rotated part absent from the gallery. With the rotated copy as a candidate, the swap shifts the preference toward it for 93% of 1,123 shapes. In natural captions a number's position and its dimension name coincide; when decoupled artificially, the model follows the position.
  • Retrieved dimensions. Excluding the correct cluster, the top retrieval's extents differ from the queried values by a median of 2.2%.

This supports sensitivity to individual overall dimensions, bound to the correct axis through their position in the caption. It does not show that the model reads the dimension words themselves.

W5 — Fusion360 exposure during development. The concern is valid. The Fusion360 gallery shapes were seen during development; only the query text of the sealed probe was new. The dev-consulted number (0.761) is not used as evidence. In the sealed probe, the caption queries passed their pre-declared bar (0.129, 108× random) and the dimension-only queries failed theirs. Our evidence for unseen geometry is the CADPrompt chain, retrained with the 190 target shapes and their 3,241 near-duplicates removed. All of it is DeepCAD-family geometry. We have no clean evaluation on a different CAD domain, and that is a real limitation.

Reviewer_QwFgOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents CALIPER, a point–text dual encoder for retrieving CAD models from geometry-only galleries using dimension-bearing descriptions. It combines gte-ModernBERT and Utonia, preserves long captions and shape scale in the corpus coordinate frame, and adapts both encoders through staged contrastive learning. Additional experiments examine text coverage, encoder adaptation, and generalization to human-written queries over held-out geometry. However, I still have several concerns, as detailed below.

Critical strengths
  1. Its main contribution appears to lie in identifying and addressing near-duplicate leakage in the original Text2CAD benchmark split, where 57.9% of test shapes have near-duplicates in the training set.

  2. The paper presents extensive experiments, i.e., on context length, encoder adaptation, and identical-text inputs, complemented by numerical perturbations, hard-pair evaluations, and failure analyses.

  3. This paper investigates an interesting and practical task, i.e., retrieving CAD models with natural-language text.

Critical weaknesses
  1. The novelty of the proposed method is limited since CALIPER mainly combines existing modules and conducts model training via standard contrastive learning and fine-tuning, without additional architectural improvements.

  2. The final comparison is not fully fair for textual input. When evaluated under the matched-text settings, the proposed CALIPER shows no clear performance advantage.

    i) As shown in Table 3, the gte-ModernBERT variant underperforms the corresponding CLIP-family encoders when given identical text: R@10 is 0.423 versus 0.472 for text truncated to 77 CLIP tokens, and 0.798 versus 0.847 for text truncated to 248 CLIP tokens.

    ii) The reported advantage of CALIPER over Long-CLIP (0.957 versus 0.909) is obtained when CALIPER reads full captions with a 2,048-token context, whereas Long-CLIP is limited to 512 CLIP tokens. This comparison therefore does not establish an advantage under matched textual input. The authors should additionally evaluate gte-ModernBERT on the same text truncated to 512 CLIP tokens for a fair comparison.

  3. Although the authors acknowledge that doubling a single numerical value barely affects retrieval performance, this still raises concerns about whether the model is able to reliably distinguish parts that differ only in a key dimension.

Questions for authors

It is suggested that the authors provide further clarification and supporting evidence to address the weaknesses discussed above.

1 comment
AuthorsOpenReview ↗

Reply to Reviewer QwFg

Thank you for the careful reading of Table 3. Retrieval numbers are cluster-credit R@10, and validation numbers are on our strict validation split.

W2 (i, ii) — No advantage under matched text. We agree, and the controls you asked for make the point sharper. We ran gte on the identical text truncated to 512 CLIP tokens, extended Long-CLIP to 2,048 tokens, and added a TULIP tower at 2,048 tokens.

Strict test, R@10:

WindowCLIP family (adapted)gte on identical text
770.4560.409
2480.8400.803
5120.9060.894 (requested control)
2,0480.929 (Long-CLIP)0.955 (CALIPER)

On validation at 2,048 tokens, TULIP reaches 0.954 against 0.957 for CALIPER (single seed). With matched text and adaptation, the gte tower has no advantage over a CLIP-family tower; it is slightly behind at 77, 248 and 512 tokens. What separates CALIPER from the zero-shot baselines of Table 1 is reading the whole caption and adapting both towers, which a CLIP-family tower does nearly as well (0.929 against 0.004 zero-shot). The submission states this in §5.3, but its headline comparison is against zero-shot baselines, and we agree that the matched comparison should have been the primary one.

W1 — Novelty. We agree. CALIPER combines existing encoders with standard contrastive training, and we do not claim architectural novelty. The new controls also show that the results depend neither on the choice of text encoder (above) nor on the cluster-aware loss, which makes no difference in a full-chain rerun (0.957 vs 0.958 with plain InfoNCE). What the paper offers is the leakage audit and split, which you identify, and controlled evidence about which factors matter for this task. We accept that the submission does not present itself that way clearly.

W3 — Parts that differ in one key dimension. We measured this directly. On 4,885 pairs of similar parts that differ by ≥10% in exactly one extent, the model picks the described part 91.0% of the time, against 50.5% with the numbers removed, and 90.1% for 10–20% differences. Two limits: accuracy is lower when the middle extent differs (83.5%), and on pairs with matched bounding boxes that differ only in internal features it is 71%, where an explicit geometric matcher reaches 86%. The model distinguishes overall dimensions reliably and internal ones only partially.

Reviewer_ekTeOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The study addresses the problem of text-to-CAD retrieval in scenarios that involve only geometry, a case in which expert queries contain exact numerical dimensions that have to be matched to the geometry at the corpus level. The authors introduce CALIPER, a point–text bi-encoder that retains the complete, lengthy captions as well as the corpus-frame scale of the point clouds, and adjusts both the text and point towers using a cluster-aware contrastive loss. They examine a significant amount of near-duplicate leakage in the standard Text2CAD split and propose a component-disjoint "strict" split; on this more rigorous evaluation benchmark, CALIPER carries out extensive ablations that separate out the effects of text coverage, geometry scale, and tower adaptation, and also provides counterfactual probes that support aggregate numeric grounding; when tested on human-written CADPrompt queries for held-out geometry, it achieves a large improvement over strong zero-shot baseline methods.

Critical strengths
  1. Carrying out systematic ablations to separate (i) the dimension text, (ii) the corpus-scale geometry, and (iii) the tower adaptation, together with carefully planned training trajectories and controls.
  2. This method adjusts both the text and the point towers within a shared space by using a cluster-aware contrastive objective, which regards near-duplicates as positives.
  3. Offers principled 'identical-text' twins which allow the text-window effects to be separated from those of the encoder; extending the CLIP-family context to 512 tokens reduces a large part of the gap, highlighting the importance of coverage.
Critical weaknesses
  1. The numeric grounding is mainly at the aggregate level since the specific values have not been shown to be linked to the named dimensions (it does not make a difference if the "length" and "width" values are swapped, a fact which may be confounded by rotational symmetry).
  2. In the training process, cluster-aware positive examples might place an emphasis on very tight geometric families; although a strict split helps to reduce leakage, an excessive dependence on near-duplicate structures could cause the learning to focus on the discrimination of individual instances rather than on broader generalisation.
  3. Although human-written evaluations are valuable, their scale is limited (with only 182 usable prompts) and they are still a long way from achieving deployment-grade accuracy; therefore, benchmarks based on a more diverse and larger scale of human queries would be needed to support the claims.
Questions for authors
  1. How great a role does cluster-aware training play in causing overspecialization among tight clusters? Have you experimented, for example, by eliminating cluster-positive averaging or by incorporating hard-negative mining in order to assess robustness?
  2. By extending the Long-CLIP to 512 tokens the gap is reduced. Have you carried out an evaluation of more recent long-context CLIP/TULIP variants exceeding 512 tokens (for example, 1k–2k) involving adaptation in order to confirm that the encoder identity contributes only marginally in the case of a comparable context?
  3. For human-written queries, can you break down the ways things can go wrong in more detail? For example, missing important numbers compared to unclear shape descriptions.
4 comments
AuthorsOpenReview ↗

Reply to Reviewer ekTe (1/2)

Thank you for the review. We ran an experiment for each of your questions and for W1. Unless marked, retrieval numbers are cluster-credit R@10 on our strict validation split.

Q2 — Long-context CLIP/TULIP beyond 512 tokens, with adaptation. We ran three controls under the identical staged recipe: Long-CLIP-L stretched to 2,048 positions, gte on the identical 512-CLIP-token strings, and a TULIP text tower run at 2,048 tokens.

Strict test, R@10 [95% CI]:

WindowCLIP family (adapted)gte on identical text
770.4560.409
2480.8400.803
5120.906 [0.899, 0.912]0.894 [0.887, 0.901]
2,0480.929 [0.923, 0.935] (Long-CLIP)0.955 [0.950, 0.960] (CALIPER)

Strict validation at 2,048 tokens, R@10 / R@1:

Long-CLIP, stretchedTULIP (relative positions)CALIPER
0.941 / 0.6710.954 / 0.7240.957 / 0.730

Your hypothesis holds. Context coverage moves R@10 by about 47 points, and the encoder family by at most 5, with a sign that changes across windows. The gap that Long-CLIP leaves at 2,048 tokens largely disappears with TULIP, which has no positional embeddings to stretch. Caveats: the TULIP run is validation-only and single-seed, its exact-uid gap is larger (0.831 vs 0.839), and at 2,048 tokens the two towers read near-identical rather than identical text. We therefore claim "within about one point", not equality.

W1 — Aggregate grounding; the length↔width swap confounded by rotation. You are right that the probe in the paper could not detect binding: the swapped caption describes a rotated part that is not in the gallery, so the original part stays on top either way. With that confound removed, binding is detected. For 1,123 shapes whose length and width differ by ≥10%, we encode the ground truth (GT) and its 90°-rotated copy (R) and compute Δ = [s(q,GT) − s(q,R)] − [s(q′,GT) − s(q′,R)] for the original caption q and the swapped caption q′.

  • Δ > 0 for 93.0% of shapes; mean +0.0246 [0.0233, 0.0260]; sign test p≈1e-215. A re-encoding placebo gives 0.000.
  • When the caption is swapped consistently (dimension values and sketch coordinates), it prefers the transformed part over the original 78% of the time for the 90°-rotated copy and 82% for the copy with the two axes exchanged.
  • In L3 captions the dimensions always appear in the same order, so the position of a number and its name coincide. When we decouple them artificially, the model follows the position: swapping only the order of the numbers gives +0.020, swapping only the words gives +0.003.

The values are therefore bound to the correct axis in natural captions, through their position in the caption rather than the dimension word. The probe covers captions with a single "length of / width of" statement (14% of validation). We will replace "not detected" in §5.4 with this result.

AuthorsOpenReview ↗

Reply to Reviewer ekTe (2/2)

Q1 / W2 — Does cluster-aware training over-specialize? We tested both variants you mention.

Removing cluster-positive averaging. The full chain rerun with plain InfoNCE, where same-cluster pairs are negatives:

stagecluster positives (paper)plain InfoNCE
heads only0.5070.505
+ point tower0.6950.698
+ text tower0.957 (exact-uid 0.839)0.958 (exact-uid 0.838)

There is no measurable difference at any stage. Cluster-aware training therefore does not cause over-specialization, and the results do not depend on it.

Hard-negative mining. We enriched batches with different-cluster shapes whose bounding boxes match within 3% and whose shapes are similar, mined on the training split with data exposure unchanged. Discrimination of bounding-box-matched pairs improves from 0.711 to 0.748 (+3.7 [+2.0, +5.4]). Retrieval drops by 0.5 R@10 points and single-extent minimal pairs by 1.5 points. It is a trade-off, not an improvement.

On generalization beyond tight families: a chain retrained with the 190 CADPrompt shapes and their 3,241 near-duplicates removed matches the original on validation (0.947 vs 0.946) and on human queries, and a new minimal-pair probe separates similar shapes from different clusters at 0.910 (0.505 without numbers). We have not tested broader CAD collections, which remains a limitation.

Q3 — Failure modes on human queries. The dominant cause is neither of the two you list but a difference in units between the prompts and our corpus. CADPrompt states dimensions in each shape's original CAD units, whereas our corpus rescales each shape so that its largest extent is 0.75; the two agree for 63 of 182 prompts.

  • Matched frame: R@10 0.810 (n = 63). Mismatched: 0.118 (n = 119).
  • Wrong-frame numbers hurt: 8.4 points below the same prompts without numbers.
  • Rescaling the lengths with the true factor gives 0.698 overall (an oracle diagnosis).
  • Missing numbers are rare (123/182 prompts state all three extents). Shape vocabulary, length and complexity are not significant in a multivariable regression.
  • Among the failures that remain after rescaling, no single cause dominates. All 15 worst cases are frame-mismatched, and two are annotation errors in CADPrompt.

§5.5 attributes the gap to information content; about half of it comes from this difference in units, and that statement needs correcting.

W3 — Scale of the human evaluation. We agree. 182 prompts, of which only 63 are in a usable frame, do not support claims about deployment, and the accuracy is far from deployment grade. A larger human-written benchmark with explicit units is needed.

Thank you for the additional experiments. While the rebuttal addresses some of my concerns, several limitations remain.

First, the new dimension-binding experiment is still restricted to length/width cases and a limited portion of the validation set. Positional association does not fully demonstrate robust semantic binding between numerical values and named dimensions.

Second, the human-query analysis reveals substantial sensitivity to coordinate-frame conventions, with only 63 of 182 queries directly matching the corpus frame. The reliance on oracle rescaling further limits practical applicability.

Finally, removing cluster-aware training produces almost identical results, questioning its actual contribution, while hard-negative mining improves pairwise discrimination at the cost of overall retrieval performance.

Overall, the additional experiments do not sufficiently establish fine-grained dimensional understanding or robust generalization to practical engineering queries.

Thank you for reading the new experiments and for the summary. One clarification: the oracle rescaling is not part of the method and not a result we claim. We used it only to confirm that the difference in units causes the failures on human queries. We agree that the remaining limitations you list are real.

Reviewer_XbYqOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose a text-based CAD retrieval system designed to improve the matching between natural-language descriptions and 3D CAD models, particularly when dimensional information is included in the queries. They propose a two-tower architecture, with one tower dedicated to text encoding and the other to point-cloud encoding. In contrast to several existing approaches, their method preserves longer textual descriptions and the scale of the 3D models in the dataset's coordinate system. Both towers are fine-tuned to learn a shared representation between text and geometry. The authors argue that preserving dimensional information is particularly important for CAD retrieval, as dimensions are often essential for identifying the correct model. They also identify data leakage caused by near-duplicate models in the standard Text2CAD dataset and propose a more restrictive evaluation split to address this issue. Finally, they report high retrieval scores on this new split, achieving a Recall@10 of 95.5% when models belonging to the same near-duplicate cluster are considered correct, and 84.8% when only the exact target model is accepted. They also report improvements over zero-shot baselines on human-written queries containing dimensions.

Critical strengths

One of the major strengths of the paper is the extensive ablation studies and experiments. They provide evidence that the model learns more than a simple matching between numerical values and bounding-box dimensions. However, these experiments do not fully establish what geometric relationships the model learns.

I noticed also a clever use of Utonia for the 3d point cloud embedding.

The authors also evaluate their approach on human-written queries, which is interesting to assess its potential use in practical applications.

Critical weaknesses

Major concerns

This paper focuses on CAD retrieval, where respecting dimensions is particularly important. The authors introduce dimensional information and provide experiments showing that their model is sensitive to numerical values. However, I do not see an experiment directly measuring whether the retrieved CAD models respect the dimensions specified in the queries. Dimensions are particularly interesting because they can be measured precisely. I think the authors are missing an opportunity to evaluate the dimensional accuracy of their retrieval results. For instance, they could measure the difference between the dimensions requested in the text and those of the retrieved models. I am therefore concerned about the practical usefulness of the proposed approach.

I am also concerned that retrieving CAD models from queries containing dimensions could partly rely on relatively simple numerical similarities. The model receives 3D points, so it can potentially infer bounding-box dimensions. The authors already provide a bounding-box baseline, but I wonder whether a more explicit geometric approach, including internal dimensions, could achieve competitive results. It would be interesting to better understand what the learned representation provides beyond matching numerical values.

Another concern is the relatively low performance on human-written queries. The authors report a Recall@10 of approximately 95% on their main evaluation and only 36% on dimension-bearing human-written queries. Although these evaluations are performed under different conditions and are not directly comparable, the performance on human-written queries remains relatively limited. This raises questions about the practical applicability of the proposed method, particularly because real users may not provide descriptions as detailed as the machine-generated captions used for training.

I am concerned that it is not sufficiently clear what the model actually learns. The authors show that numerical information improves retrieval performance, but they do not demonstrate that the model learns the relationships between dimensions or their relative proportions. In particular, the experiments in Section 5.4 suggest that the model is not very sensitive to individual dimensional changes. This raises questions about whether it learns meaningful geometric relationships or mainly statistical correlations between numerical values and shapes.

I think more controlled experiments are needed to better understand what the model learns and to demonstrate its usefulness for practical CAD retrieval.

Minor concerns

The related work section reads more like a collection of approaches than a structured discussion of previous research. I think a more conventional presentation, with paragraphs connecting and comparing the different approaches, would make this section more interesting and easier to follow.

Questions for authors
  1. Have you considered designing an experiment to measure the difference between the dimensions specified in the queries and the actual dimensions of the retrieved CAD models? Is this type of evaluation outside the scope of your paper?
  2. Does CALIPER outperform a simple numerical or geometric baseline, particularly when the queried dimensions refer to local geometric features rather than global bounding-box dimensions?
  3. In Section 5.4, the experiments suggest that the model is sensitive to global numerical changes but much less sensitive to modifications of individual dimensions. Do you think this indicates that the model mainly learns a global numerical representation rather than the relationships between dimensions? How could you explain this behavior?
  4. How does CALIPER compare with Text-to-CAD Retrieval: a Strong Baseline (2026)? Have you considered including this approach in your experimental comparisons?
2 comments
AuthorsOpenReview ↗

Reply to Reviewer XbYq (1/2)

Thank you for the detailed review. We ran the experiments you suggested and report them below, including where the outcome does not favour us. Unless marked, retrieval numbers are cluster-credit R@10 on our strict validation split.

Q1 / W1 — Do retrieved models respect the queried dimensions? You are right that the submission did not measure this. We now do, as the per-axis relative error between the queried dimensions and the retrieved shape's bounding-box extents (n = 7,760). The table gives the fraction of axes within 5% for the top-1 retrieval. The ground-truth cluster is excluded, so it describes the best alternative, not the correct answer.

Top-1 retrieved byall axeslargest axis excluded
CALIPER0.578 (median error 2.2%)0.435
query with numbers stripped0.2140.056
random0.2870.049
GT-bbox oracle0.9550.958

Retrieved shapes follow the queried dimensions well above chance. Two limits should be stated:

  • CALIPER is not a precise dimension matcher. On captions whose overall dimensions can be parsed, a matcher that compares those numbers to bounding boxes is more precise on this metric (0.768 vs 0.576), although on the same captions it finds the right part less often (R@10 0.712 vs 0.966).
  • On human CADPrompt prompts the effect is weak. The same measurement on the 182 prompts, with the model retrained without the CADPrompt shapes, is below. Excluding the ground-truth cluster, prompts with dimensions are not distinguishable from random.
Human prompts (n = 182), fraction of axes within 5%top-1top-1, GT cluster excluded (median error)
prompt with dimensions0.399 [0.350, 0.454]0.317 [0.271, 0.359] (26%)
prompt without dimensions0.3040.289 (34%)
random0.3110.311 [0.286, 0.335] (36%)

Q2 / W2 — A simple numerical or geometric baseline, including local features? We built an explicit geometry-only matcher. It parses every number in the caption and matches it against values measured on the point cloud: bounding-box extents, planar levels (thicknesses, step heights), and fitted circle radii and centres.

R@1R@10bbox-matched pairs (as in Table 5)
bbox shortcut (paper)0.1570.2490.546
explicit geometric matcher0.4270.6290.862
CALIPER0.7290.9570.711
CALIPER + matcher (score fusion)0.7980.9640.835

All four rows of the last column use the same 2,000 pairs. Table 5 reports 0.713 for CALIPER from an earlier draw of pairs. The bbox shortcut can parse overall dimensions from only a third of the captions; on that third its R@10 is 0.712 (CALIPER 0.966), and on the rest it is at chance.

The answer to your question is mixed. In full-gallery retrieval the learned model is far ahead (0.957 vs 0.629). On pairs with matched bounding boxes that differ only in internal features, the explicit matcher is clearly better (0.862 vs 0.711). The learned representation therefore uses internal dimensions only partially, and the two are complementary. Training with mined bounding-box-matched hard negatives narrows this gap (0.711 → 0.748) but does not close it, and costs 0.5 R@10 points.

One further reference: matching the numbers against each shape's CAD construction parameters reaches R@10 0.967, above CALIPER. It requires the construction history, which a geometry-only gallery lacks, and the captions were generated from those parameters.

AuthorsOpenReview ↗

Reply to Reviewer XbYq (2/2)

W3 — Human-written queries (0.36). Your concern is justified. Analysing the failures showed that the main cause is a difference in units between the CADPrompt prompts and our corpus. The shapes are the same DeepCAD shapes, but CADPrompt states dimensions in each shape's original CAD units, whereas the corpus rescales every shape so that its largest extent is 0.75. The two scales agree for 63 of the 182 prompts. For another 63 the stated numbers are twice the corpus values, and for 56 the ratio is something else.

nR@10, prompts with dimensions
prompt frame = corpus frame630.810 [0.714, 0.905]
prompt frame ≠ corpus frame1190.118

Rescaling only the length values with the true factor raises R@10 from 0.357 to 0.698. This is an oracle diagnosis, not a result we claim. Missing numbers are rare (123/182 prompts state all three extents), and shape vocabulary, prompt length and complexity are not significant in a multivariable regression. Two consequences follow:

  • §5.5 attributes the gap to the information content of the query. That explanation needs correcting: about half of the gap comes from this difference in units.
  • Your practical concern stands. The corpus carries no absolute units, so a deployed system needs a mapping from user units to the corpus frame, which we did not address. Even where the frames match, the evidence rests on 63 prompts.

Q3 / W4 — Individual dimensions and relationships (§5.4). Doubling one value barely changed the rank for two reasons. An L3 caption contains dozens of numbers, most of them construction parameters. And retrieval is relative to the gallery: if no shape matches the edited value, the original stays on top. That probe was therefore uninformative, and we ran two better ones.

  • Minimal pairs. On 4,885 pairs of similar shapes that differ by ≥10% in exactly one extent, forced-choice accuracy is 0.910 [0.905, 0.916], against 0.505 with the numbers stripped. It is 0.901 for differences of 10–20%, and weaker when the middle extent differs (0.835) than the smallest (0.964).
  • Role binding. The length↔width swap in the paper was confounded, because the swapped caption describes a rotated part that is not in the gallery. With the rotated copy added as a candidate, the swap shifts the preference toward it for 93% of 1,123 shapes (p≈1e-215). L3 captions always list the dimensions in the same order, so a number's position and its name coincide; when we decouple them artificially, the model follows the position rather than the word.

So the model resolves a 10–20% change in one overall dimension and, in natural captions, binds it to the correct axis. We have no evidence that it reads the dimension words themselves, or that it represents relations between dimensions beyond this. Note also that the corpus fixes the largest extent at 0.75 for about 89% of shapes, so most stated dimensions are in effect proportions of that extent.

Q4 — "Text-to-CAD Retrieval: a Strong Baseline" (2026). This is the paper we cite as IMMA (arXiv 2605.05572); we should have used its title. We could not include it in our main comparisons for two reasons. Its code is not public, so we cannot run it on our component-disjoint split or with long captions. And it encodes each gallery item's parametric command sequence, which a geometry-only gallery does not provide. Instead, App F reproduces IMMA's own protocol: the official split, short abstract captions for training and evaluation, and exact-uid credit. A CALIPER model trained on abstract captions was evaluated once on the official test set after pre-registration.

Official test, abstract captions, exact-uidR@1R@10median rankgallery input
IMMA (reported)0.0970.36125point cloud + command sequence
CALIPER0.0850.37920point cloud

The two are close. IMMA is better at R@1, which we attribute to its sequence input: many shapes share identical abstract captions, and the construction sequence can tell them apart while geometry alone often cannot. This is a comparison against self-reported numbers, not a controlled one. Our default model, which is not trained on abstract captions, is much weaker in this setting.

Minor — related work. We agree and will rewrite it as a comparative discussion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.