- 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
-
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.
-
The proposed method is lightweight and easy to follow.
-
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:
-
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.
-
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:
- 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
- Have the authors evaluated whether the reconstructed meshes also preserve realistic overall body dimensions and species-specific size variations?
3 comments
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+ | 0.594 | 14.5 | 26.7 | 709.0 | 1.16 | pass |
| Head fine-tuned, no pathway (new) | 0.638 | 22.2 | 26.7 | 710.0 | 1.13 | pass |
| 2D keypoints only (Table 3) | 0.655 | 37.9 | 119.1 | 706.1 | 1.69 | fail |
| Ours | 0.643 | 30.2 | 10.3 | 707.5 | 0.93 | pass |
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.
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
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.
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.
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.