acceptodds
Under review as a conference paper at ICLR 2027

Species-Aware Morphometric Priors for Avian 3D Reconstruction

Abstract

Reconstructing a bird in 3D from one photograph remains hard. There are 11,000 species of extreme variety, almost no 3D shape data, and no absolute scale in the image. Annotations mark one point per part, which says where a part is but not how long it is. Current methods regress the parameters of a bird model built from a few species, and they recover pose and silhouette well. Yet the part proportions they produce barely change from one species to another and do not match those measured for the species. Our starting observation is a difference between birds and humans. Humans remain one species however skin, height or limb length varies. Birds share the same joints, yet how far the bill, tarsus, tail and wing extend from those joints differs by species and is likely what tells one species from another. Ornithology measures exactly these lengths, and we bring them into 3D reconstruction from AVONET, an open database of more than 90,000 measured individuals across 11,000 species. Since a photograph has no scale, we follow the AVONET definitions and express each length as a ratio to wing length. Since these ratios vary little within a species, the species-average ratios serve as a species-aware morphometric prior. A morphometric pathway realizes it without disturbing the pretrained body model. A proportion head infers per-part length factors from the features of a pretrained transformer reconstructor, and an axial deformation operator inside the bird model applies them before skinning. Because the body model is left intact, pose and camera adapt to the new part lengths and species morphology is added without relearning the shape space. No photograph carries a measurement: for real images training uses species averages alone, and inference uses one photograph. On 49 held-out species the median error of the tail, bill and tarsus proportions against the measured species values falls from 21–31% without the prior to about 10%. On 237 species never seen in training it falls from 25–61% to 11–18%, while 2D keypoint accuracy improves and the 3D error is preserved. Long-legged species such as herons, cranes and storks, absent from the training photographs, remain difficult, and extending the prior to them is the next task.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
3D bird reconstruction, single-image 3D reconstruction, shape prior, morphometrics, birds

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_THJ5OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the limited ability of existing single-image avian 3D reconstruction methods to capture species-specific morphological proportions. The authors introduce species-aware morphometric priors derived from AVONET, using measured tail, bill, and tarsus lengths relative to wing length as training supervision. Experiments on held-out and unseen bird species show good performance than previous method especially on 2D keypoint accuracy.

Critical strengths
  1. The paper identifies a clear limitation of existing avian reconstruction methods that is not adequately captured by conventional keypoint and silhouette metrics. Incorporating real biological measurements as species-level supervision is well motivated.

  2. The proposed method is lightweight and easy to follow.

  3. The experiment design is relatively comprehensive, including species-level generalization, comparisons with alternative adaptation strategies, and component ablations. Results on unseen species are convincing.

Critical weaknesses

Major comments:

  1. The experimental comparison is primarily limited to AniMer+, while other relevant avian 3D reconstruction methods, such as DOVE and AVES, are only discussed in the related work. As authors’ method is based on AniMer+, including those model-free methods as baselines would make this paper better shaped.

  2. The main evaluation design measures agreement between predicted mesh proportions and species-average measurements from AVONET. While useful, this does not directly establish that the reconstructed geometry accurately represents the individual bird in the input image. Additional experiments on the DOVE dataset, which provides ground-truth 3D meshes, would further strengthen the evaluation of 3D reconstruction accuracy.

I would consider raising my rating if the authors could provide convincing results addressing my concerns.

Minor comments:

  1. The proposed model improves PCK over AniMer+, but the 2D-only variant in Table 3 achieves an even higher PCK. This suggests that the 2D improvement may partly result from additional keypoint supervision or pose/camera fine-tuning rather than the species-aware morphometric prior itself.
Questions for authors
  1. Have the authors evaluated whether the reconstructed meshes also preserve realistic overall body dimensions and species-specific size variations?
3 comments
AuthorsOpenReview ↗

Response to Reviewer THJ5 (1/2): W1 baselines, W2 evaluation on 3D scans

We thank Reviewer THJ5 for the careful review and for recognising that the limitation we identify is not captured by conventional keypoint and silhouette metrics.

[W1. DOVE and AVES as baselines]

The paper studies a pathway that attaches to a template-based reconstructor, so one reconstructor serves as the host, and AniMer+ is the strongest public one for birds. We add both methods the reviewer names as baselines, and they answer the question behind the weakness: whether the proportion gap is a property of AniMer+ or of template-based bird reconstruction in general.

AVES (Wang et al., 2021). AVES is the parametric model that AniMer+ regresses, and its original paper fits it by optimisation, so we add that route as its own row, in a setting deliberately favourable to the baseline: the full shape space (15 shape coefficients, all 24 bone scales, pose, camera) is optimised for 300 Adam steps against ground-truth (GT) keypoints and masks on CUB-49 (mask IoU 0.90).

Table R1. CUB-49, tail / bill / tarsus. Err.: median relative error of the species proportions (%). r: correlation with the measured species ratios. CV: spread of the predicted species medians (%; measured 42 / 31 / 35). Tail target scaled by k as in Table 1.

Err. ↓r ↑CV
AniMer+31.3 / 21.0 / 27.9−0.32 / 0.22 / −0.176.6 / 1.6 / 2.4
AVES fit (GT keypoints + masks)33.6 / 26.5 / 23.0−0.23 / 0.45 / −0.344.8 / 3.8 / 3.1
Ours (image only)8.9 / 11.1 / 11.00.87 / 0.79 / 0.9028 / 23 / 28

Even with GT 2D evidence and the whole shape space free, the errors stay at the AniMer+ level and the species spread is a tenth of the measured one. This is the central observation of the paper: a keypoint marks where a part is, not how long it is, so 2D evidence does not determine part length, whichever reconstructor consumes it and however well it fits. The gap is therefore not an artefact of AniMer+. The species measurements supply what 2D fitting cannot, and the pathway does so without any annotation at test time, in one forward pass rather than several seconds of optimisation per image.

DOVE (Wu et al., 2023). DOVE learns its own template without fixed vertex semantics, so a part ruler would depend on landmarks of our choosing. We therefore ran its released model on the benchmark where it defines the protocol, its own 3D Toy Bird scans (W2).

[W2. 3D accuracy on ground-truth meshes]

We evaluated on 3D Toy Bird (Wu et al., 2023; 23 scanned toys, 345 photographs) with DOVE's released evaluation script, applied identically to every method. No toy image is used in training.

Table R2. Chamfer ↓ on 3D Toy Bird (cm, bird width 10 cm; mean ± s.d. over images). Ours closer: images (of 345) / toys (of 23), Wilcoxon p.

Chamfer ↓Ours closer ↑
DOVE, released model (published: 1.51 ± 0.89)1.43 ± 0.93257 / 18 (p < 10⁻²⁸)
AVES fit to GT masks (IoU 0.94)1.44 ± 0.77294 / 22 (p < 10⁻³⁸)
AniMer+1.16 ± 0.84285 / 22 (p < 10⁻³⁹)
Ours (3 seeds: 0.92 / 0.93 / 0.93)0.93 ± 0.63

Our protocol reproduces DOVE's published value within its spread, so the rows are comparable. Our reconstructions are a fifth closer to the real scans than AniMer+'s and a third closer than DOVE's or the AVES fit's. The three baselines, in contrast, are close to one another: AniMer+ is closer than DOVE on 193 of 345 images, and the AVES fit and DOVE are indistinguishable (p = 0.36). Ours comes clearly closer to the scans than this group, and against AniMer+, which shares our template and regressor, the difference is the species prior and not fine-tuning (control in Table R3). Correcting proportions with species statistics therefore brings the mesh closer to the true 3D shape of the individual object, which is what the species-average protocol could not show.

We will add Tables R1 and R2 with their protocols to the paper.

AuthorsOpenReview ↗

Response to Reviewer THJ5 (2/2): minor point and question

[Minor. PCK of the 2D-only variant]

The submission reports the 2D numbers as a check that the prior costs nothing in 2D accuracy, not as a contribution of the prior. To separate how much of the PCK increase comes from fine-tuning the head, we trained a control with exactly the same losses and schedule but without the pathway (3 seeds).

Table R3. [email protected]: CUB test set. Tail-tip PCK (%) and mean proportion error (%): CUB-49. PA-MPVPE: synthetic test set. Chamfer: 3D Toy Bird (cm). Mesh check: the gate of Section 4.1.

[email protected] ↑Tail-tip PCK ↑Prop. err. ↓PA-MPVPE ↓Chamfer ↓Mesh check
AniMer+0.59414.526.7709.01.16pass
Head fine-tuned, no pathway (new)0.63822.226.7710.01.13pass
2D keypoints only (Table 3)0.65537.9119.1706.11.69fail
Ours0.64330.210.3707.50.93pass

On PCK alone, fine-tuning the head accounts for most of the increase (0.594 → 0.638), and our model keeps all of it and adds a little (0.643), the highest of every row that passes the mesh gate. On the columns that measure geometry, the control stays where AniMer+ was: its proportion error is unchanged at 26.7 % against our 10.3 %, and its real-scan Chamfer is 1.13 against our 0.93 (ours is closer on 286 of 345 images). Fine-tuning moves the keypoints and nothing else; what moves the proportions and the real 3D shape is the prior. PCK does not see part length, which is why Section 4.2 treats it as a check that the proportions are not bought with the standard metrics, not as the measure of what the prior does. The 2D-only variant makes the same point from the other side: the highest PCK of the table together with the largest proportion error, a failed mesh check and the worst real-scan error.

[Q. Overall body dimensions and size variation]

Two things are being asked, and the answer to the second is yes. Relative body build is observable, and we tested it with a trait never used in training or evaluation, AVONET body mass^(1/3) / wing length against mesh volume^(1/3) / wing length: the correlation across species goes from 0.415 (AniMer+) to 0.445 (ours) on CUB-49 and from 0.412 to 0.423 on NABirds-237. Body build is inherited from the pretrained reconstructor, which the pathway leaves untouched by design (Section 3.3), and it is preserved, with a slightly higher correlation and no degradation. The pathway can also take body build on directly: a body axis with an AVONET mass-derived target would be a fourth part under Eq. 5. Absolute size is a different matter: a photograph has no scale, so no single-image method recovers it.

We hope that the two new baselines, the evaluation on real 3D scans and the fine-tuning control address the concerns raised, and we would be grateful if the reviewer would reconsider the rating in their light.

AuthorsOpenReview ↗

Follow-up to Reviewer THJ5

Dear Reviewer THJ5,

We sincerely thank you for reconsidering our paper and for raising your rating. Your suggestions on the AVES and DOVE baselines and on the evaluation on real 3D scans have clearly strengthened the paper, and these results will be included in the revised version.

If any questions remain, or if there is anything further we could clarify or add that would further strengthen your assessment of the paper, we would be more than happy to address it before the discussion period closes.

Thank you again for your constructive engagement throughout the review process.

Best regards, The Authors

Reviewer_yUU5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper shows that an avian reconstruction model can recover pose and silhouette while producing part proportions that vary little across species. The authors use species-average ratios from AVONET to supervise a morphometric pathway that adjusts tail, bill, and tarsus lengths before skinning, with pose and camera predictions adapted accordingly. The method requires no individual measurements for real training photographs and no species label at inference. Evaluation covers proportion error on held-out and additional unseen species, 2D keypoints, synthetic 3D accuracy, and mesh integrity.

Critical strengths

1.The paper identifies a weakness that standard reconstruction metrics can miss. Section 4.2 shows substantial variation in measured proportions across species but little corresponding variation in AniMer+ reconstructions. This is a useful finding: good keypoint or silhouette scores do not necessarily imply plausible species morphology. Using physical measurements for both supervision and evaluation gives the work a clear purpose. 2.The improvement in species-level proportion prediction is well supported. In Table 2, the average of the three part-wise errors falls from 26.7% to 10.3% on CUB-49 and from 39.5% to 14.0% on NABirds-237. Keypoint accuracy improves, while synthetic 3D error stays close to the baseline. Multiple seeds, comparisons with alternative formulations, and analyses of unseen species strengthen the evidence. The reported failures and compression of extreme proportions also help define the method’s limits.

Critical weaknesses

1.Passing the mesh gate does not establish reliability for individual reconstructions. Sections 4.1–4.3 use this check to support mesh quality, but Table 8 (p. 21) applies it to medians across images. In the rest-mesh diagnostic on 678 held-out photographs, Table 10 (p. 22) reports a gate failure on 20.2% of images and at least one flipped triangle on 29.1% for the proposed pathway. The authors disclose these defects, and the flips usually involve few faces. Please report the per-image failure rates there and explain what these local defects mean for using the reconstructed meshes. 2.The synthetic 3D comparison with post-reconstruction stretching mixes architectural and supervision differences. Section 4.3 and Table 3 use this baseline’s higher PA-MPVPE to support the internal pathway. On p. 20, however, the baseline is described as being trained on species-average ratios, while the proposed proportion head also receives synthetic ratio supervision. Table 6 (p. 19) shows that removing this supervision raises the proposed method’s PA-MPVPE from 707.5 to 804.4. The baseline’s error of 827.5 therefore does not, by itself, show that the gap comes from placing the deformation inside the reconstructor. Please clarify whether the baseline receives the same synthetic supervision. If it does not, a matched comparison would help; alternatively, the conclusion should be limited to the advantage of the complete training setup.

Questions for authors

1.How much do the local degeneracies and flips in Table 10 affect the usability of the meshes? Is there evidence that tighter tail-deformation bounds reduce these defects while retaining most of the proportion improvements? 2.Does the ratio predictor in the post-reconstruction baseline receive synthetic ratio supervision? If not, how much of the PA-MPVPE gap remains when it is trained with the same supervision?

2 comments
AuthorsOpenReview ↗

Response to Reviewer yUU5

We thank Reviewer yUU5 for the careful review and for recognising that good keypoint or silhouette scores do not by themselves imply plausible species morphology.

[W1, Q1. Per-image failure rates, their source, and what the defects mean for use]

Per-image rates. As suggested, the per-image rates of Table 10 move to Table 2, and Sections 4.1–4.3 will say which checks apply to medians and which per photograph.

Where the defects come from. Attributing each flipped face to the deformation coefficient that produces it, the tarsus factor alone produces 99.6 % of the flipped faces and the tail less than 0.1 %. This also answers the question about the tail bounds: clamping the tail factor at −0.70, or even −0.50, leaves the per-image rates unchanged, because the tail is not the source.

Table R5. CUB-49 / NABirds-237, 3 seeds. Gate fail, Flipped: photographs failing at least one gate condition / with at least one flipped face (%). Flips: median per affected photograph; worst. Err.: proportion error on CUB-49, tail / bill / tarsus (%).

Gate fail ↓Flipped ↓Flips ↓Err. ↓[email protected] ↑PA-MPVPE ↓
AniMer+0.0 / 0.00.0 / 0.00; 0 / 031.3 / 21.0 / 27.90.594709.0
Ours (paper)20.2 / 23.229.1 / 37.76; 83 / 1078.9 / 11.1 / 11.00.643707.5
Ours, tarsus bound [0.7, 1.4], retrained6.6 / 10.022.5 / 32.82; 11 / 128.3 / 10.5 / 11.80.643706.9

Retraining with the tarsus bound cuts the per-image gate failures to a third and the worst photograph from 83 flipped faces to 11, while proportions, PCK and 3D error stay within seed variation. We adopt this bound as the recommended setting; the main tables of the paper stand as they are.

What the defects mean for use. The roughly 140 CUB-49 photographs that fail the gate are harder for every method: AniMer+, which has no defects, loses about ten PCK points on them, and on those same photographs our model is still about five points above AniMer+. A flipped face therefore marks a difficult photograph rather than a reconstruction that is worse than the baseline.

[W2, Q2. Supervision of the post-hoc baseline and the matched comparison]

The predictor of the Table 3 baseline was trained on species averages only, while our proportion head also receives Lsyn, so the two rows differ in supervision as well as in placement. We ran the matched comparison the reviewer asks for: the same predictor with the same synthetic ratio loss (3 seeds) and, since our model also fine-tunes the head, the same predictor on a head fine-tuned without the pathway.

Table R7. Mean proportion error (%) and tail-tip PCK (%): CUB-49. [email protected]: CUB test set. PA-MPVPE: synthetic test set. Chamfer: 3D Toy Bird (cm; AniMer+ 1.16).

Prop. err. ↓[email protected] ↑Tail-tip PCK ↑PA-MPVPE ↓Chamfer ↓
Post-hoc stretching (Table 3)9.10.59420.0827.5–
Post-hoc + Lsyn (new)9.40.59419.0710.20.76
Post-hoc + Lsyn, fine-tuned head (new)9.40.63834.1711.60.75
Ours without Lsyn (Table 6)10.20.64230.5804.4–
Ours10.30.64330.2707.50.93
  • Lsyn keeps the 3D shape accurate wherever the prior is applied. It lowers PA-MPVPE to the AniMer+ level for both placements, so the PA-MPVPE gap of Table 3 belongs to Lsyn.
  • The 2D gap of Table 3 belongs to head fine-tuning, not to placement. With the head fine-tuned, post-hoc stretching reaches our PCK and a higher tail-tip PCK, and its meshes pass the per-image checks at the same rates as ours.
  • On the real scans the post-hoc rows are closer than ours. All are far below AniMer+ (1.16), so the real-scan gain belongs to the prior and to Lsyn, which the rows share. The remaining difference is the tail: on these short-tailed toys Chamfer falls as the tail is shortened, and the post-hoc predictor shortens tails further (mean factor −0.60 against our −0.25). The toys carry no species label, so the correct tail length cannot be judged there; on CUB-49, where it can, the rows reach the same proportion error.

We will therefore revise Section 4.3 as the reviewer suggests: it will state these attributions, add the matched rows to Table 3, and present the in-reconstructor pathway as one of two equivalent realisations. Ours needs a single forward pass and trains the proportion head jointly with pose and camera, whereas the post-hoc route solves the deformation for each photograph at test time. The matched comparison strengthens the main claim: the species prior with Lsyn improves proportions and real 3D shape however it is attached.

We hope that the per-image analysis and the matched comparison resolve the two weaknesses, and we would be grateful if the reviewer would take them into account in the final assessment.

AuthorsOpenReview ↗

Follow-up to Reviewer yUU5

Dear Reviewer yUU5,

We sincerely thank you once again for the time and effort you have devoted to reviewing our paper.

As the author–reviewer discussion period approaches its conclusion, we hope that our responses, in particular the per-image mesh analysis with the tarsus bound and the matched comparison with the post-hoc baseline, have addressed the points you raised. If any questions remain, or if there is anything further you would like us to clarify or provide, we would be more than happy to do so before the discussion closes. If you feel that your concerns have been sufficiently resolved, we would be grateful if you would consider updating your rating accordingly.

We truly appreciate your thoughtful feedback and consideration throughout the review process.

Best regards, The Authors

Reviewer_YP3FOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses the problem of inaccurate species-specific body proportions in single-image 3D bird reconstruction. The method leverages ornithological measurements to introduce species-aware morphometric priors, improving the reconstruction of anatomical proportions across different bird species.

Critical strengths
  1. Leveraging ornithological measurements to improve 3D bird reconstruction is a well-motivated idea. Drawing inspiration from anthropometric priors in human body reconstruction and adapting this concept to species-specific bird morphology is an interesting approach.
  2. The experimental evaluation is well presented. Since the method builds upon AniMer+, the differences between the baseline and the approach are clearly demonstrated both quantitatively and qualitatively, including in the appendix.
Critical weaknesses

The method mainly builds on existing techniques, including anatomical measurement-based supervision. It also models limited anatomical variation. Although it improves specific bill, tarsus, and tail proportions, it does not address anatomical variations across the entire bird body.

Questions for authors
  1. Since the main evaluation relies on species-average morphometric measurements, how well does the method capture anatomical variations between individuals of the same species?

  2. What is the reason for not including other body parts or measurements, such as overall leg length or body length? What were the main challenges?

  3. Could your approach also improve 3D reconstruction for other animals, such as dogs? Is there anything specific to birds or the AVES model that makes your approach particularly effective here?

3 comments
AuthorsOpenReview ↗

Response to Reviewer YP3F

We thank Reviewer YP3F for the careful review and the positive assessment.

[W. Novelty and the scope of the anatomical variation]

Components. The contribution lies in three findings rather than in the components. Each component exists on its own (virtual measurements in SHAPY, length scales in AVES, the AVONET database), and we deliberately kept each simple so that other reconstructors can adopt the pathway.

  1. A failure invisible to standard metrics. Part proportions of current bird reconstructors barely vary across species while measured ones vary by 35–43 % (Section 4.2). A new experiment that fits the full AVES shape space to ground-truth keypoints and masks shows that this holds even then: these length scales are not determined by image evidence.
  2. Supervision from species averages, not from measurements of the photographed bird. The measurement-based supervision the reviewer refers to, SHAPY in the human case, uses measurements of the person in the image. We use species averages, so no photographed bird is ever measured, and the result generalises to 237 species never seen in training.
  3. These lengths lie outside the bird shape space. Pushing the shape parameters toward them breaks the mesh on almost every photograph (Table 10), which is why a separate pathway is needed.

Scope. The three parts are those that AVONET measures for essentially all of the roughly 11,000 species, and the lengths in which species differ as a matter of function rather than ornament (Section 1). The pathway itself is not limited to them (Q2).

[Q1. Variation between individuals of the same species]

The paper targets the variation between species, because that is where the proportions of birds differ: 93–95 % of the variance of the AVONET log ratios lies between species, and within a species the ratios vary by about 6 % (Section 3.2). Two checks nevertheless ask whether the individual in the photograph is still read.

  • On synthetic birds, the correlation between predicted and ground-truth individual ratios is on par with AniMer+ (Appendix H), and it is Lsyn that keeps it so.
  • On real photographs no dataset measures the photographed bird, so the only individual difference both labelled in photographs (NABirds) and measured in the field (AVONET) is sex. Over 39 species, the predicted male–female difference of the tail ratio correlates with the measured one at r = 0.55 (permutation p = 0.046) against 0.09 for AniMer+, rising to 0.66 on the 22 species whose measured difference exceeds 1 % and to 0.92 on the 7 above 2 % (AniMer+ 0.04 and 0.17). Photograph by photograph, a male falls on the measured side of a female of the same species with AUC 0.57 against 0.47. No sex information is used in training.

A model trained only on species averages therefore keeps large individual differences and moves them in the measured direction, whereas the base reconstructor shows no such signal.

[Q2. Other body parts and measurements]

The pathway takes any part that has a root joint, an axis and a weight (Eq. 5), so the limit was not the method but the measurements and the photographs.

  • Measurements. AVONET records four lengths per species: wing (our ruler), tail, bill and tarsus. There is no tibia, femur or body length in AVONET.
  • Photographs. The tibia is covered by body feathers in most species, and body length depends on posture.
  • The way forward. The skeletal trait database of Weeks et al. (2025) now reports tibiotarsus and femur lengths for thousands of species. That is the route to the whole leg and to the long-legged waterbirds of Section 4.4.

[Q3. Other animals]

What makes the approach effective here is a property of birds: their part proportions are a species trait to an unusual degree. Within a species the ratios vary by about 6 %, while across the roughly 11,000 species they vary by 35–43 % (Section 3.2), so a table of species measurements carries almost all of the variation, and AVONET supplies that table for nearly every species. In humans, by contrast, a different build changes the segment lengths together and the ratios between them stay similar (Figure 1), so a body model can treat them as nearly fixed.

To transfer the pathway, an animal therefore needs three things: a template with part axes, a table of part lengths per group, and proportions that vary mostly between groups. For quadrupeds the template (SMAL) and the measurement tables both exist. For dogs the groups are breeds, whose proportions differ much as bird species do, and breed standards together with multi-breed measurement sets (Sutter et al., 2008: 27 body measurements over 109 breeds) can play the role of AVONET. Thank you for the suggestion; we will pursue it as follow-up work.

We hope these answers clarify what is new in the paper and how far the method reaches, and we would be glad to expand on any of them during the discussion period.

AuthorsOpenReview ↗

Follow-up to Reviewer YP3F

Dear Reviewer YP3F,

We sincerely thank you once again for the time and effort you have devoted to reviewing our paper, and for your positive assessment.

As the author–reviewer discussion period approaches its conclusion, we hope that our responses on the novelty of the work, within-species variation (including the new sexual-dimorphism analysis over 39 species), other body parts and other animals have addressed your questions. If any questions remain, or if there is anything further you would like us to clarify or provide, we would be more than happy to do so before the discussion closes. If you feel that your concerns have been sufficiently resolved, we would be grateful if you would consider updating your rating accordingly.

We truly appreciate your thoughtful feedback and consideration throughout the review process.

Best regards, The Authors

Reviewer_YP3FOpenReview ↗

I will keep my positive rating

Dear authors, thank you for your detailed response! I also learned much more about birds thanks to your paper. I'll keep my positive rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank the three reviewers for their careful and constructive reviews, and we are encouraged that all three recognise the problem the paper identifies and find the species-level improvement well supported. We ran eight new experiments. Each is described in the thread of the reviewer who raised it, and all will be in the revised PDF with changes in blue.

  • Reviewer THJ5. An AVES optimisation baseline fitted to ground-truth keypoints and masks (proportion error stays at the AniMer+ level); evaluation on real 3D scans, DOVE's 3D Toy Bird (Chamfer ↓: DOVE 1.43, AVES fit 1.44, AniMer+ 1.16, ours 0.93); a head fine-tuning control without the pathway; relative body size (preserved).
  • Reviewer yUU5. A per-image mesh analysis (flipped faces come from the tarsus factor, not the tail); a tarsus bound adopted as the recommended setting at unchanged accuracy; a post-hoc baseline with matched supervision and fine-tuning (Section 4.3 is revised accordingly).
  • Reviewer YP3F. Within-species variation via sexual dimorphism over 39 species (the predicted male–female tail difference correlates with the measured one, r = 0.55 vs 0.09).

Paired statistics for the main result: our tail / bill / tarsus error is lower than AniMer+'s on 45 / 37 / 39 of the 49 held-out species and on 208 / 181 / 163 of the 237 unseen species (Wilcoxon signed-rank, all p < 10⁻⁵).

We hope these results address the concerns raised, and we would be grateful if the reviewers would take them into account in their final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.