acceptodds
Under review as a conference paper at ICLR 2027

CARE3R: Contact-Aware Feed-Forward 3D Reconstruction for Surgical Endoscopy

Abstract

Three-dimensional reconstruction from endoscopic video supports surgical planning, intraoperative navigation, augmented visualization, and robotic assistance. Yet a reconstruction can be accurate pixel by pixel and still place the instrument tip at the wrong distance from the tissue. We show that this instrument–tissue clearance is dissociated from standard geometry metrics: three variants of our model trained with different instrument supervision have near-identical depth, pose, and reconstruction metrics, including the same tissue depth error on our stereo validation clips, yet their contact AUCs range from 0.74 to 0.93; and fine-tuning the prediction heads of Depth Anything 3 on our training data reduces its C3VD depth error sevenfold without improving contact discrimination. To make clearance measurable and learnable, we introduce ESD-Contact, a multimodal ex-vivo endoscopic submucosal dissection (ESD) dataset with synchronized stereo and side views, verified contact/hover labels, and resistance recordings for a subset of clips, together with a contact benchmark that scores contact/hover discrimination from primary-view RGB. We further propose Contact-Aware Supervision (CAS), which supervises the instrument with targets taken from independent measurements rather than from the model's own prediction: a stereo-referenced depth profile along the instrument, and perpendicular tip-to-tissue targets that a triangle construction derives from the side view. CARE3R, trained with CAS jointly across public and in-house datasets, predicts metric-scale depth, pointmaps, camera poses, and instrument segmentation from RGB frames in one feed-forward pass. On the validation clips of ESD-Contact, it reaches a contact AUC of 0.933, compared with at most 0.806 for seven general and surgical 3D baselines, while remaining competitive on conventional depth, pose, reconstruction, and segmentation benchmarks. Code and model weights will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Endoscopic Reconstruction, Monocular Depth Estimation, 3D Feedforward Reconstruction, Medical Image Analysis

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

The authors argue that instrument-tissue clearance is not captured by standard geometry metrics. They proposed CARE3R, which is a VGGT-style feed-forward model that predicts metric depth, pointmaps, poses and instrument masks from endoscopic RGB. They contribute ESD-Contact, a contact-vs-hover AUC benchmark, and Contact-Aware Supervision. On four validation clips CARE3R reaches AUC 0.933 versus ≤0.806 for seven zero-shot baselines.

Critical strengths

A real, clinically motivated gap, addressed with an original data design. Using a side view to derive perpendicular clearance targets, with resistance recordings to verify labels on a subset, makes a millimetre-scale quantity learnable and measurable.

Dense FS supervision alone gives AUC 0.822, a self-referential profile hurts (0.743, the tip sinks), a stereo-referenced profile gives 0.883, and contact/hover terms add 0.050, positive on every clip.

Critical weaknesses

The headline claim rests on four validation clips that were also used to select loss settings, with a single seed and no confidence intervals. I appreciate that held-out surgical data are scarce and the authors acknowledge this, but a leave-session-out split , bootstrap CIs, and separating tuning from evaluation are feasible. CARE3R exceeding the stereo teacher (0.933 vs 0.892) suggests it may exploit visual contact cues.

Conventional metrics are measured on tissue or other datasets, and no tip-region depth error is reported.

Questions for authors

Could the authors report depth error in the tip region against FS, and the hover clearance error in mm against the side-view measurement? How does the model behave on “near” frames, which are currently excluded?

Would it be possible to evaluate with a leave-session-out split, multiple seeds, and bootstrap CIs? Arm D suggests variation of about 0.005 AUC and 1 mm, so how large is the uncertainty on the four-clip AUC?

How can the AUC (0.933) exceed the stereo teacher (0.892) that supplies the targets? Does the model rely on visual cues rather than geometry?

3 comments

We thank the reviewer for the valuable comments. All three requests, namely bootstrap CIs, a session-disjoint check and seed variation, are now run, and we have revised the manuscript where appropriate. The four-clip AUC carries real uncertainty (±0.03\pm 0.03); the paired contrast does not (+0.050+0.050, positive in every replicate); and it survives removing temporal context and same-session clips.

Re W1 / Q2 (a): how large is the uncertainty

A two-level bootstrap, resampling clips then frames, gives [0.902,0.963][0.902, 0.963] around the four-clip AUC of 0.9330.933, and [+0.030,+0.071][+0.030, +0.071] for the +0.050+0.050 paired contrast. Per-clip AUCs are 0.966, 0.956, 0.900 and 0.909, so the interval reflects clip-to-clip variation.

Re Q3 / W1: whether the gain comes from visual cues rather than geometry, and why we can exceed FS

In one line: we train no contact classifier, so an appearance cue could only enter the score by making the predicted depth wrong; and FS is not a ceiling on this metric, for reasons a model with temporal context can exploit. First, there is no contact output. The network predicts metric depth and point maps, camera poses and an instrument mask; contact is never a prediction target. Every method, ours and all seven baselines, is scored by the same readout on its own depth map: the median depth in a patch at the annotated tip minus that of a surrounding tissue ring. AUC ranks that scalar. Second, the supervision is that same distance, not a label. On a contact frame the loss drives the tip's signed perpendicular distance to a plane fitted to the surrounding tissue ring toward zero; on a hover frame it drives that distance toward the side-view measurement. The plane comes from a detached copy of the model's own ring, so the tip is never supervised by its own depth. Reducing it requires moving the tip to the right place in 3D. Third, the paired contrast isolates the supervision:

SupervisionContact termsw=8w{=}8w=1w{=}1
Dense FS (stage 1)none of our modules0.8220.815
REL + FS profileabsent0.8830.874
CARE3R (ours)present0.9330.921
Why we exceed FS. FS scores 0.8920.892, and that is not a ceiling on this metric. The labels are judged in the side view and checked against the resistance trace, and hover clearance comes from the side-view distance, so FS is a frame-by-frame depth estimate scored against a label it did not produce. Each frame is also reconstructed on its own, so its error is independent across frames and a model seeing several frames can average it out. Finally the readout is a local tip-to-tissue distance, which the contact terms target directly, whereas dense depth supervision spends capacity on every pixel. The time column tests this: removing temporal context costs ours 0.0120.012 and leaves the gain at +0.047+0.047, while stage 1, seeing only dense FS depth, never reaches FS at either window and sits 0.1060.106 below ours.

Re W1 / Q2 (b): same-session clips and seeds

In one line: removing every same-session clip costs no more than re-finetuning on an unrelated session, and three seeds agree to ±0.001\pm 0.001. The experiment acts on the contact stage from the reported model and tests whether the supervision depends on same-session data; it is not a run from base, which we label. A control keeping all ten clips under the same recipe divides out that cost (−0.007-0.007, CI [−0.016,+0.001][-0.016, +0.001]). Values are relative to it:

Armheld-out clipother three clips
SD0714−0.012 [−0.020,−0.005]-0.012\ [-0.020, -0.005]−0.012 [−0.018,−0.005]-0.012\ [-0.018, -0.005]
SD0731, seed 42−0.004 [−0.014,+0.006]-0.004\ [-0.014, +0.006]−0.005 [−0.011,+0.002]-0.005\ [-0.011, +0.002]
SD0731, seed 43−0.011 [−0.021,−0.003]-0.011\ [-0.021, -0.003]−0.002 [−0.006,+0.003]-0.002\ [-0.006, +0.003]
SD0731, seed 44+0.006 [−0.006,+0.018]+0.006\ [-0.006, +0.018]−0.005 [−0.010,−0.000]-0.005\ [-0.010, -0.000]
Two observations agree: 20260630/clip_1, the only clip whose session contributes no training data, is the highest-scoring at 0.966; and every session-disjoint arm stays above the contact-free model. The three seeds give four-clip means of 0.9216, 0.9220 and 0.9234, a standard deviation of 0.0010, since every component we introduce lives in the contact stage. No checkpoint was selected on the evaluation clips.

Re W2 / Q1: why not report tip region error against FS, and near frames

In one line: we do not report it because its reference is FS, an estimate rather than ground truth, whereas the contact state is annotated in a second view and used as the criterion. An error at the tip would be measured against FS depth, itself a prediction and least reliable at instrument edges. Closer agreement would not show that the geometry is recovered, so methods should not be ranked by it. The contact state is what we can hold fixed: annotated in the side view, and on the three clips with a resistance channel the two agree on 99.9% of frames. Near frames are excluded because the transition is ambiguous; only 21 exist in the four clips.

The response says FS depth is “least reliable at instrument edges.” But Appendix A.2 says: “FS is reliable near the instrument but noisier on distant, weakly textured tissue.” The paper also uses the Spearman correlation between predicted and FS tip–tissue gaps as supporting evidence. So it’s hard to see why the authors won’t report tip-region error against the same reference. I also have a concern about the two-level bootstrap. The outer level has only four clips, so the interval mostly reflects variation across those four clips. It may understate the real uncertainty. More evaluation clips would make this stronger in future work. Overall, the paper tackles a clinically relevant gap with an original data design. I’m keeping my weak accept rating.

AuthorsOpenReview ↗

Clarification: where FS is reliable, and where it is not

We thank the reviewer for the careful re-reading and for keeping the paper under consideration. Both points are fair. Our wording in the previous response described where FS is weakest in a way that did not match Appendix A.2, and the bootstrap interval is indeed limited by having four clusters. We make the picture clearer on both below.

Clarification: where FS is reliable, and where it is not

Our earlier sentence said FS is "least reliable at instrument edges". What we meant is the small region around the instrument tip and the tissue immediately around it, not the instrument surface, which FS measures well: it is close to the camera, large in the image and strongly textured. FS is noisier on tissue, most of all on distant, weakly textured tissue. We will correct the wording.

The gap is a difference between two depth readings. One is taken on the instrument at the tip, where FS is accurate. The other is taken on the tissue just outside the instrument mask, where FS is weaker. In hover the two readings are only a few millimetres apart, while the tissue itself is tens of millimetres from the camera. A small relative error on the tissue reading is therefore a large relative error on the gap, and that is the reading that decides the score.

FS's own score on this benchmark says the same thing empirically. Run through the identical readout against the side-view labels, FS reaches an AUC of 0.892 on the stereo clips, not 1.000. If its depth resolved the gap well enough to serve as the reference for it, its own readout would be close to perfect.

This does not make FS a poor supervision signal. As dense supervision it is clearly better than nothing: it is dense, metrically scaled, and correct in ordering over most of the frame. The claim is narrower. FS is good enough to supervise against and to correlate against, and not precise enough on tissue to serve as the absolute reference for a millimetre gap.

Why ρ\rho is reported and a tip-region error is not

Both use FS, but they ask different things of it. Spearman ρ\rho asks only that the predicted and FS gaps rise and fall together across frames, and a smooth error or an offset on the tissue term does not break that ordering. An absolute error asks that FS be accurate at the very location where we have just said it is weakest. The primary metric does not depend on FS either: AUC is scored against the side-view contact labels, and the per-frame alignment to FS changes it by less than 0.002 for every method.

The bootstrap

We agree. With four clips the outer level has three degrees of freedom, so the interval is coarse and probably optimistic.

Three checks bound the claim independently of it. The statistic we rely on is the paired contrast of +0.050+0.050, which is positive on all four clips individually (+0.023+0.023 to +0.083+0.083), so its sign does not rest on the interval. Dropping any single clip leaves a mean of 0.9220.922 to 0.9440.944, so no one clip carries the result, and even the lowest of these is 0.1160.116 above the strongest baseline (0.806). And the two checks that do not use the four clips agree: three seeds give 0.92160.9216, 0.92200.9220 and 0.92340.9234 (SD 0.00100.0010), and removing every same-session clip costs at most 0.0120.012 while every arm stays above the contact-free model.

We agree that more evaluation clips is the right fix, and we state the four-clip scope as a limitation. We will also add this clarification to the manuscript, so that the reliability of FS on the instrument and on tissue is stated once and consistently.

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

The pape introduces ESD-Contact, an ex-vivo dataset and benchmark, and CARE3R, which uses independently measured instrument depth and clearance for supervision. It shows that accurate 3D reconstruction does not necessarily capture instrument–tissue contact. The method improves contact discrimination while remaining competitive on standard reconstruction tasks.

Critical strengths
  • The paper introduces both a dataset and a reconstruction method, providing useful contributions to instrument–tissue contact estimation.
  • The main idea is clear and easy to understand.
Critical weaknesses
  • The dataset contains only 14 videos, with four held out for validation. This small scale limits the evidence for generalization across patients, instruments, and surgical conditions.
  • The benchmark uses annotated instrument tips and masks, hence it does not evaluate the full pipeline under realistic detection and segmentation errors.
  • Per-frame alignment using a stereo-derived tissue reference provides additional information during evaluation. Performance without this alignment should also be reported to clarify.
  • The paper treats contact as inferred from visual geometry, which cannot establish contact force or distinguish light touching from tissue compression. The resistance recordings could provide stronger physical validation.
  • The results is evaluated only on 4 in-house datasets, leaving transfer to independent datasets and unverified.
Questions for authors

Overall, I see the paper is interesting, but still missing some key details: 1. the scale/effect of small dataset. 2. why instrument-tissue contact has no "force" - which is more needed in practical application (even though it's not easy to estimate).

3 comments

We thank the reviewer for the constructive assessment. We address each point below and have revised the manuscript where appropriate.

Re W1: dataset scale

In one line: 8,241 is what survives curation of roughly 50,000 raw frames and the curation is what makes the set annotation-dense rather than large. The raw recordings contain 50,193 frames, of which 8,241 are retained (16.4%) after out-of-focus rejection, near-duplicate removal, annotation-coverage checking, and manual quality control. Scale is modest but not an outlier, since widely used surgical 3D datasets such as C3VD and SCARED have about 10k and 29k frames. What differs is annotation density: masks, tissue labels, keypoints in both views, an interaction state, and metric depth on every retained frame. This is a first release of the benchmark, and we state the 14-clip scope as a limitation rather than a generalisation result.

Re W2: annotated tips and masks

In one line: this is what makes the benchmark usable by geometry-only models, at the cost of measuring geometry rather than the full pipeline. Any method predicting depth or point maps can be scored without a detection or segmentation head, which is what made the seven-model comparison possible; the numbers are therefore an upper bound on end-to-end performance. We will release the evaluation code so that fuller outputs can be scored without reimplementation.

Re W3: alignment, and the FS-supervised fine-tune

In one line: there may be a misunderstanding, because we do report a baseline fine-tuned with FS depth, DAv3 (FT), in the main contact table, and strengthening that supervision makes contact discrimination worse rather than better. Depth Anything 3 (Large) is initialised from its official weights and its prediction heads are fine-tuned on our training clips with FoundationStereo (FS) depth as the target, without contact supervision:

ArmC3VD AbsRelContact AUC
DAv3 (zero-shot)0.6330.800
DAv3 (FT, FS depth)0.0890.777
A sevenfold reduction in depth error leaves contact AUC slightly below zero-shot, so FS depth supervision is not a route to the contact gain.
Our inference never uses FS: the input is monocular RGB from the primary view only. The per-frame tissue-median alignment is an evaluation-time normalisation applied identically to every method, because several baselines predict relative depth and have no metric scale; without it the comparison would measure scale convention rather than geometry. Removing it moves the four-clip AUC by less than 0.0020.002 for every method, at either window (ours 0.93290.9329 to 0.93320.9332 and 0.92130.9213 to 0.92200.9220), and we report the unaligned column.

Re W4 / Q2: contact and force

In one line: this paper targets the relative geometry of instrument and tissue, which the endoscopic view can observe, while force is a separate quantity that would need a sensor at the instrument tip, which this setup does not have. The benchmark scores whether the tip touches the tissue and how far it lies from it, a relation that prior surgical 3D estimation leaves unmodelled and that a feed-forward reconstruction from the primary view can measure. A gap near zero tells us that the tip has touched the tissue, but not how hard it presses. A light touch and a firm press can give the same gap: the tissue under the tip deforms, and the reference surface moves with it. The force also depends on how stiff and large the contact is, and the gap records neither. It need not pass through the tip either, since the blade and shaft carry force. Measuring it needs such a sensor. The resistance channel we record is a binary electrical-contact signal falling from the megaohm range to tens of kilohms; it verifies the interaction state and does not measure compression, so it cannot stand in for a force measurement. Combining that geometry with force is a natural next step, for which the synchronised two-view recordings we release provide the geometric half.

Re W5: why contact is in-house

In one line: the relative geometry of instrument and tissue is a gap in current surgical 3D models, which is why we built this benchmark, and no public dataset provides the synchronised two-view recording it requires. Existing surgical datasets target tissue reconstruction, depth, pose or segmentation, but none carries a per-frame, metric statement about the instrument and tissue relationship, which the endoscopic view alone cannot observe. Contact is therefore evaluated in-house because no public equivalent exists. Contact is the only task evaluated on a single in-house dataset. Depth, pose and reconstruction are evaluated on C3VD, SCARED and SERV-CT, and segmentation on EndoVis18 and the Kvasir test split, none of which we collected.

We initially took the one-response-per-reviewer limit to be firm. Having found that additional messages are accepted, we add the following, which our earlier reply had to compress to stay within the length limit.

Additional detail on Re W1: what curation removed, and what the clips contain

The 50,193 recorded frames fall to 8,241 in four passes: out-of-focus rejection, near-duplicate removal, a coverage check that keeps only spans with enough tissue landmarks visible in both views to anchor the side-view construction, and a manual pass over every propagated label. Retention varies from 7% to 44% per clip because the recordings were not scripted and the knife left the usable field of view at different rates.

The four validation clips illustrate what the retained frames carry:

ClipFramesHoverContactNearUnknownView
0630-1645516115140stereo
0714-271857713074stereo
0828-178156921200stereo
0731-3712302388022monocular
Total2,8561,9648452126

Of these, 1,963 hover and 843 contact frames are scored; two contact and one hover frame on the monocular clip lack a valid readout. The state of every frame is judged in the side view, and near frames (21 across the four clips) and unknown frames (26, mostly contacts with the tip occluded in the side view) are excluded because the transition is ambiguous by construction.

Additional detail on Re W3: the alignment result in full

Modelw=8w{=}8 alignedw=1w{=}1 alignedw=8w{=}8 unalignedw=1w{=}1 unaligned
CARE3R (ours)0.93290.92130.93320.9220
w/o contact and hover terms0.88300.87390.88310.8740
Pi30.80570.81130.80490.8103
DAv30.79970.82180.80020.8229

Alignment moves no model by more than 0.001 at either window. Without temporal context ours drops by 0.012 and the two strongest baselines rise slightly, so our margin is not an artefact of the window. The paired contact gain is +0.050 at eight frames and +0.047 at one, positive on every clip at both settings.

One clarification on the row labelled w/o contact and hover terms. It removes only the contact, hover and floor terms. REL, the FS-referenced blade profile and the raised normal weight are all retained, and it is the same stage-2 model trained from the same stage-1 checkpoint. Its 0.8830.883 is therefore not a second dense-FS baseline. Against stage 1 (0.8220.822), which has none of these terms, the three remaining terms add 0.0610.061 on their own. They help because they regularize the instrument's depth directly: removing the instrument from the dense loss, constraining the blade profile against FS, and strengthening the normal term stop dense supervision from dragging the tip toward the tissue. What they do not do is tell the model when the tip is touching, which is what the contact and hover terms add.

Additional detail on Re W4: what the resistance channel does establish

The LCR channel is a binary electrical-contact signal, falling from the megaohm range during hover to tens of kilohms during contact. It is not a force measurement, and we do not treat it as one. Its role is to cross-check the labels: on the three sensor clips the de-bounced trace agrees with the side-view state on 99.9% of frames (1,061 of 1,062), whereas the raw trace agrees on 83.1%, the difference being residual mucus that bridges the electrodes after lift-off. That is why the trace proposes states and an annotator decides.

AuthorsOpenReview ↗

Look forward to your post-rebuttal feedback

Dear Reviewer EQih,

Thank you again for taking the time to review our paper and share your feedback. We hope our responses have addressed your concerns.

As the discussion period comes to an end, we wanted to check if you have any further questions or comments. We would be happy to clarify anything that is still unclear.

Thank you for your time and thoughtful suggestions.

Best regards, The Authors

Reviewer_rXZYOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies instrument–tissue contact geometry in surgical endoscopic reconstruction. It argues that standard frame-level depth and reconstruction metrics can miss local errors near the instrument tip. The authors introduce ESD-CONTACT, which combines primary and side views, stereo-derived depth, and resistance recordings for some clips to provide interaction labels and local geometric annotations. CARE3R adds contact-aware supervision to a feed-forward multi-view reconstruction model. Contact and hover losses guide tip placement relative to a local tissue plane. A stereo-referenced instrument depth profile also constrains the instrument’s shape. On four validation clips, CARE3R achieves a mean contact AUC of 0.933, compared with 0.806 for the strongest external baseline. Internal comparisons show that contact discrimination can improve even when standard geometry metrics barely change. The method also remains competitive on depth, pose, reconstruction, and segmentation tasks. Overall, the results mainly support better contact/hover discrimination, rather than accurate absolute clearance estimates.

Critical strengths
  1. A well-motivated problem formulation supported by a valuable dataset and targeted evaluation The paper addresses an important failure that average geometric errors can hide: whether the reconstructed instrument tip is correctly positioned relative to the tissue. The authors demonstrate this with model variants that achieve similar standard geometry scores but very different contact AUCs. ESD-CONTACT combines side-view observations, interaction labels, and local geometric annotations to study this issue directly. Using annotated tips and masks to isolate geometric errors also fits the research goal. The clear problem definition, dedicated data collection, and targeted validation are key reasons for my positive assessment.

  2. Geometrically motivated supervision with supporting internal comparisons The method improves the reconstructed geometry rather than just adding a contact classifier. Contact and hover labels guide tip placement, while the shaft-centered FS profile constrains the instrument’s shape. This makes it harder to satisfy the clearance loss by deforming only the tip. Applying stop-gradient to the tissue plane also clarifies where the loss directly drives updates.

Critical weaknesses
  1. The geometric interpretation and reliability of the hover targets require clearer qualification The main text says that hover targets come from independent measurements. However, the appendix explains that these distances are estimated from side-view DA3 depth and scaled using a known shaft length. About 45% of the selected anchors also have two valid positive roots, and the implementation always chooses the larger one. These targets are therefore independent of the model being optimized, but they are not necessarily accurate or uniquely determined physical distances. The authors are suggested to clarify how the primary view, stereo reference, and side view are registered and combined.

  2. Validation-driven configuration selection and benchmark contact-label construction need further clarification The appendix states that the four contact validation clips were used to choose contact-loss settings. The main contact results are then reported on those same clips. The authors clarify that these clips were not used to select the final checkpoint, but it is still unclear which hyperparameters or design choices were guided by validation performance. Separating choices based on training data from those based on validation results would help clarify what the reported findings support. Interaction labels are also central to the benchmark. The authors should clearly define what counts as contact and explain how manual annotations and LCR resistance measurements are combined to assign the final labels.

Questions for authors
  1. Multi-view registration and integration: How are correspondences and consistent scales established across the primary view, stereo reference, and side view? Does the triangle construction require explicit inter-camera extrinsics, or only cross-view keypoint correspondences and side-view distances?

  2. Contact labels and LCR measurements: What operational criteria distinguish contact, near, and hover in the benchmark? For clips with LCR recordings, how are initial labels derived from resistance values, and what rules guide subsequent manual corrections?

  3. Configuration selection: Which hyperparameters and design decisions were informed by performance on the four validation clips, and which were determined solely from training-set statistics or predefined geometric principles? Please distinguish configuration selection from checkpoint selection.

3 comments

We thank the reviewer for a careful reading, and for pressing on the two points where our presentation was thinner than our evidence: the physical status of the hover targets, and the boundary between validating a design and selecting on the metric. We have revised the manuscript where appropriate.

Re W1 / Q1: hover targets and multi-view registration

In one line: the targets are independent of the model but not uniquely determined, and the construction needs no inter-camera extrinsics. We agree and will correct the main text. Registration uses no relative pose between the two cameras. The primary view gives the tip ray through the intrinsics, which needs no tip depth. A local tissue plane fitted to FS depth on a two-ring pattern gives the metric scale. The side view gives one point-to-point distance per annotated tissue landmark, from DA3 relative depth, undistorted and scaled by the known shaft length ∣AC‾∣=9.20 mm|\overline{AC}| = 9.20\,\mathrm{mm}. The views are temporally synchronised and corresponded through shared tissue landmarks rather than dense matching, and frames are rejected when the reconstructed tip is inconsistent with the marked shaft points (∣AB‾∣=5.25|\overline{AB}| = 5.25, ∣BC‾∣=3.95 mm|\overline{BC}| = 3.95\,\mathrm{mm}). With ρ\rho the side-view distance, rr the primary-view ray-to-anchor distance and θ\theta the angle between them, d⊥2=ρ2−r2sin⁡2θd_{\perp}^{2} = \rho^{2} - r^{2}\sin^{2}\theta admits two positive roots when ρ\rho is short enough. The distance alone cannot separate them, so the implementation keeps the larger root. This affects 38% of accepted anchors and 45% of those used by the hover term, and the main text will describe the target as the largest clearance consistent with the side view.

Re W2 / Q2: contact labels and LCR

In one line: the side-view annotation is the label of record, and the sensor proposes while an annotator decides. States are hover, near, contact and unknown, judged in the side view. Near frames (41 in corpus, 21 evaluated) and unknown frames (131 in corpus, mostly tip occluded) are excluded from supervision and from the benchmark. On the three LCR clips (2,334 frames) resistance falls from MΩ\mathrm{M}\Omega during hover to tens of kΩ\mathrm{k}\Omega during contact; an automatic pass derives initial states, then an annotator verifies every frame, because residual mucus bridges the electrodes after lift-off. The de-bounced trace agrees with the side-view labels on 99.9% of frames (1061/1062).

Re W2 / Q3: configuration selection

In one line: no checkpoint was selected, every weight was fixed before evaluation, and the variants we retained are the ones that do not raise contact AUC.

CategoryDecisionBasis
Checkpointlast stage-2 checkpoint (19)fixed rule; no selection on any clip
Physical priorsτabs=−0.5 mm\tau_{\text{abs}} = -0.5\,\mathrm{mm}, τnorm=−0.01\tau_{\text{norm}} = -0.01fixed before evaluation
Training statisticshover floor ϕ=1.04 mm\phi = 1.04\,\mathrm{mm}, the 90th percentile of contact distancestraining clips only
Gradient calibrationcontact, REL, PROFILE and NORM weightstraining data only
Design variantsstage-2 arm; FLOOR, PROFILE, NORMcompared on the four clips
We will state the last row explicitly rather than describe the clips as held out. Three facts bound what it can inflate. The paired contrast is within recipe, with the same data, schedule and checkpoint index, so +0.050+0.050 cannot depend on which checkpoint was kept. The retained settings are not the ones that raise contact AUC: across the cumulative ablation arms A to D it stays within 0.9260.926 to 0.9280.928, with per-clip changes of at most 0.0090.009 that change sign, and they were kept for gap magnitude and surface quality. Arm D, a full re-train with a different initialisation, schedule and seed, lands at 0.9280.928 against our 0.9330.933, which is the spread between independent runs rather than a selection gain.
Reviewer_rXZYOpenReview ↗

Comment by Reviewer rXZY

Thank you for the detailed response and clarifications. The authors have satisfactorily addressed my concerns.

AuthorsOpenReview ↗

Thank you for your encouraging feedback

Dear Reviewer rXZY,

We sincerely thank you for your positive feedback and are very pleased that our clarifications and additional experiments have resolved your concerns. We greatly appreciate your support and positive recommendation, and your constructive comments have been invaluable in improving our paper. Thank you again for your time and thoughtful review!

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.