PaperlayerView exact revision on arXiv

Source-linked AI summary

TrackFish3D: Self-Supervised 3D Tracking of Schooling Fish from Multi-view Videos

Patt Phurtivilai, Zhiyang Dou, Yifan Wu, Kinfung Chu, Yuan Liu, Lei Yang, Wenping Wang, Taku Komura

arXiv:2609.38347v1cs.CV

TL;DR

Dense 3D tracking of schooling fish is difficult because of occlusions, visual similarity, and scarce identity annotations. TrackFish3D learns spatial and temporal associations from calibrated multi-view geometry, achieving 95.8% MOTA on SynFish versus 87.7% for the strongest baseline and 81.1% on 3D-ZeF versus 77.4% for the best geometric baseline.

  • Problem

    Dense 3D trajectories are needed to quantify collective fish behavior, but multi-view tracking is challenged by frequent occlusions, visually similar individuals, and scarce identity annotations.

    Source: §1 · pp. 1–2

  • Method

    TrackFish3D uses calibrated multi-view geometry to learn cross-view correspondence and temporal identity consistency without identity labels, appearance re-identification, or ground-truth trajectories.

    Source: §1 · p. 2

  • Results

    95.8% MOTA with SAM3 versus 87.7% for the strongest baseline on SynFish, and 81.1% MOTA versus 77.4% for the best geometric baseline on 3D-ZeF.

    Source: §4.3 · p. 7, §4.4 · p. 8

  • Takeaways & Limitations

    TrackFish3D's recovered trajectories are reported as accurate enough for downstream abnormal-behaviour detection, and the formulation also transfers to real-world pigeons.

    Source: §5 · p. 10

    TrackFish3D assumes known camera intrinsics and extrinsics and requires per-frame, per-camera 2D detections as input.

    Source: §3 · p. 3

Abstract

from arXiv · show

Quantifying collective fish behavior requires accurate trajectories, yet multi-view 3D tracking remains challenging due to frequent occlusions, visually similar individuals, and the long-standing scarcity of identity annotations. We present TrackFish3D, a geometry-driven self-supervised framework for dense multi-camera 3D tracking of schooling fish. Instead of relying on appearance-based re-identification or manually annotated identities, TrackFish3D turns calibrated multi-view geometry into supervision: triangulation and reprojection consistency provide pseudo-associations, while a geometric encoder and global association transformer learn all-to-all cross-view correspondence within each frame. To make these associations identity-aware, TrackFish3D introduces a self-supervised contrastive objective that separates co-visible individuals in the embedding space, together with a temporal predictor that preserves identities and bridges short occlusions across frames. The resulting model is trained once on unlabeled footage and applied directly to unseen test videos, requiring no cross-view identity labels, temporal annotations, 3D ground truth, appearance features, or test-time optimization. On our benchmark, TrackFish3D improves 3D Multi-Object Tracking Accuracy from 87.7% for the strongest baseline to 95.8%. On the 3D-ZeF zebrafish benchmark, it achieves 81.1% MOTA, compared with 77.4% for the best geometric baseline. TrackFish3D also generalizes beyond fish, achieving strong results on real-world bird tracking.

1 Introduction

Dense 3D trajectories are needed to study naturalistic schooling, but fish are visually similar, frequently occlude one another, and lack identity annotations. TrackFish3D learns cross-view and temporal identity consistency from calibrated geometry, and the paper introduces SynFish for evaluation.

Motivation

  • Dense 3D trajectories are needed to understand naturalistic schooling, but visually similar fish, frequent occlusions, and scarce supervision make spatial and temporal association difficult.Spatial association fuses multi-view observations into 3D positions; temporal association links them into identity-consistent trajectories.Source: §1 · pp. 1–2

Existing approaches

  • Existing trackers leave gaps: single-view methods cannot fuse views, multi-view association omits temporal continuity, and joint trackers require identity supervision.Self-supervised multi-view methods still rely on visual features.Source: §1 · p. 2

Approach

  • TrackFish3D learns cross-view correspondence and temporal identity consistency from calibrated geometry, without identity labels, appearance re-identification, or ground-truth trajectories.Triangulation and reprojection consistency provide supervision; learned embeddings separate co-visible individuals and temporal predictions propagate identities through occlusions.Source: §1 · p. 2

Results

  • TrackFish3D outperforms all methods on SynFish and 3D-ZeF, achieving higher 3D tracking accuracy and fewer identity switches.Evaluation uses a unified protocol that converts representative baselines into end-to-end 3D tracking pipelines.Source: §1 · p. 2

Contributions

  • SynFish is a realistic Unreal Engine 5 benchmark with dense ground-truth 3D trajectories for visually homogeneous schools across group sizes, occlusion levels, and camera configurations.The paper also evaluates on the real-world 3D-ZeF20 zebrafish benchmark [47].Source: §1 · p. 2

2 Related Work

Prior tracking approaches solve only parts of dense multi-camera 3D tracking or depend on identity supervision and appearance cues. TrackFish3D targets the gap by jointly learning geometric cross-view association and temporal identity consistency for visually homogeneous animals.

Multi-view association from 2D detections

  • Classical multi-view association triangulates matched detections but becomes fragile with occlusions, missed detections, and localization noise.Learned MVA methods may reduce annotation needs, but still solve only per-frame spatial association and rely on visual features.Source: §2 · p. 3

Single-view tracking followed by 3D lifting

  • Single-view trackers model temporal association within one camera, but cross-camera association of fragmented tracklets reintroduces ambiguity for visually interchangeable animals.Their per-view tracks must still be associated across cameras to obtain 3D trajectories.Source: §2 · p. 3

Joint multi-view multi-object tracking

  • Joint multi-view trackers address spatial and temporal association, but typically require identity supervision and discriminative appearance features.ADA-Track++ assumes 3D object detections and targets autonomous-driving scenes with visually distinct objects.Source: §2 · p. 3

TrackFish3D

  • TrackFish3D jointly learns geometric cross-view association and temporal identity consistency without identity labels, appearance re-identification, or ground-truth 3D trajectories.It targets the underexplored intersection of geometry-only identity association and self-supervised temporal 3D tracking for visually homogeneous animals.Source: §2 · p. 3

3 Method

TrackFish3D reconstructs 3D trajectories from calibrated per-camera detections by jointly learning cross-view association and temporal identity consistency without identity annotations or ground-truth 3D trajectories.

3.1 Geometric encoder

  • A lightweight MLP encodes normalized detection geometry, camera identity, confidence, and a back-projected world ray into a 128-dimensional appearance-agnostic embedding.Inputs include centroid and bounding-box measurements; the representation encodes geometric constraints without visual appearance.Source: §3.1 · p. 4

3.2 Association transformer

  • A global self-attention transformer processes all detections jointly, using scene-wide cross-view context to estimate correspondence probabilities between detections from different cameras.Rotary positional embeddings encode normalized centroid positions to help distinguish nearby detections in dense scenes.Source: §3.2 · pp. 4–5

3.3 Geometric pseudo-supervision

  • Triangulation and reprojection consistency label cross-camera pairs as matches when the triangulated point falls inside both detections’ bounding boxes.Positive-pair confidence is based on average reprojection error, modulating supervision strength without changing the binary pseudo-labels.Source: §3.3 · p. 5

3.4 Self-supervised association loss

  • A confidence-weighted binary cross-entropy loss trains cross-camera correspondence predictions against geometry-derived pseudo-labels, separately normalizing positive and negative terms.This prevents negative pairs from dominating when positive pairs are scarce.Source: §3.4 · p. 5

3.5 Multi-view grouping & triangulation

  • Learned pairwise scores and embedding agreement support Hungarian matching, transitive merging across cameras, and triangulation; reprojection validation retains consistent groups.Group-member embeddings define contrastive positives, while each group embedding serves as an identity token for temporal linking.Source: §3.5 · p. 5

3.6 Self-supervised contrastive loss

  • An InfoNCE-style contrastive loss treats detections in the same multi-view group as positives and all other detections as negatives, separating identities within a frame.The objective pulls same-fish detections together across views and pushes different identities apart.Source: §3.6 · p. 6

3.7 Temporal linking and self-supervised temporal loss

  • A temporal predictor uses matched group embeddings across consecutive frames, with a weight-normalized mean squared error encouraging identity-consistent embeddings over time.The temporal objective improves linking under occlusion.Source: §3.7 · p. 6

3.8 Training and Inference

  • Training combines association, contrastive, and temporal objectives; inference reconstructs 3D hypotheses and links them over time, bridging missing tracks with a learned predictor.Inference requires no identity labels, appearance re-identification, or privileged scene information.Source: §3.8 · p. 6

4 Experimental Results

TrackFish3D is evaluated on synthetic fish scenes, real-world zebrafish, and a pigeon subset, alongside ablations of its training losses, appearance features, and association transformer.

4.1 Datasets and difficulty characterisation

  • SynFish has 14 scenes with 4–16 visually similar fish; training uses eight scenes, with six held-out scenes for testing, including an unseen camera rig.The real-world 3D-ZeF split trains on zebra1 and zebra3 and tests on zebra2 and zebra4; difficulty is also characterized by overlap measures.Source: §4.1 · pp. 6–7

4.2 Baselines, Protocol, and Metrics

  • Evaluation compares 13 complete 3D tracking pipelines using identical detections, calibrations, and post-processing, with reconstructed trajectories scored by standard 3D MOT metrics.The benchmark includes supervised and self-supervised multi-view approaches, single-view trackers lifted to 3D, and multi-camera MOT methods.Source: §4.2 · p. 7

4.3 Results on SynFish test set

  • 96.6% MOTA with YOLO26+SAM2 and 95.8% with SAM3; the strongest baseline, 3D-SORT, reaches 87.7% on held-out SynFish scenes.TrackFish3D remains above 86% MOTA on dense 16-fish scenes, while baselines drop sharply beyond 8 fish.Source: §4.3 · p. 7, Table 2 · p. 8, Figure 5 · p. 8

4.5 Cross-Species Generalization to Real-World Pigeons

  • On Pigeon10, TrackFish3D achieves 91.9% MOTA, 94.5% recall, and fragmentation of 4 without bird identity labels.The subset mainly contains ground-walking pigeons with limited flight, so it is preliminary evidence beyond fish, not validation on sustained 3D flight.Source: §4.5 · p. 8

4.4 Results on 3D-ZeF test set

  • 81.1% MOTA on 3D-ZeF exceeds the Hungarian geometric baseline's 77.4%; TrackFish3D also reduces fragmentation to 30.5 and ID switches to 4.0.3D-ZeF contains only five fish, making per-frame cross-view association relatively sparse.Source: §4.4 · p. 8, Table 3 · p. 9

4.6 Ablation Studies

  • Removing contrastive loss drops MOTA from 97.7% to 36.8%, while removing temporal loss lowers it to 86.0% and raises identity switches from 0.5 to 2.0.Removing both auxiliary losses yields 36.9% MOTA, nearly identical to removing contrastive loss alone.Source: Table 4 · p. 9, §4.6 · p. 9
  • Removing global self-attention reduces MOTA by 17.4 percentage points, lowers mostly-tracked trajectories from 95.8% to 64.2%, and increases fragmentation from 6.5 to 16.3.Adding DINOv3 features yields 97.9% MOTA versus 97.7% for geometry alone, with lower MTBF and doubled identity switches.Source: §4.6 · pp. 9–10, Table 5 · p. 10

5 Conclusion

TrackFish3D uses calibrated geometry to learn cross-view association and temporal identity linking without identity labels, appearance-based re-identification, or ground-truth 3D trajectories. It reports strong results on SynFish and 3D-ZeF while noting dependencies on calibration, multi-view visibility, and reliable detections.

5 Conclusion

  • 96.6% MOTA on held-out SynFish scenes and 81.1% on real 3D-ZeF zebrafish are reported, alongside improved identity stability over strong baselines.The method also transfers to real-world pigeons, and recovered trajectories support proof-of-principle abnormal-behaviour detection.Source: §5 · p. 10
  • The method depends on calibration, multi-view visibility, and reliable detections; future work could improve robustness to calibration errors and missed detections.The conclusion also calls for dense real-world multi-camera animal datasets with ground-truth 3D trajectories.Source: §5 · p. 10

Appendix Contents

  • Appendix experiments cover 30-fish scaling, pseudo-label noise, training-time calibration perturbations, runtime, and cross-species generalisation to pigeons.The contents also list architecture, training, inference, dataset, calibration, per-scene, ablation, and downstream-analysis sections.Source: Appendix Contents · p. 15

A TrackFish3D Training Algorithm

The training algorithm takes calibrated scenes and 2D detections and learns an encoder, association transformer, and temporal predictor. It combines geometric pseudo-supervision with association, contrastive, and temporal objectives.

A TrackFish3D Training Algorithm

  • Training takes scenes with calibrated cameras and 2D detections, and outputs an encoder, association transformer, and temporal predictor.A sliding-window track memory is maintained across consecutive frame pairs.Source: A TrackFish3D Training Algorithm · p. 16
  • For each frame pair, the algorithm encodes detections, predicts cross-camera associations, and obtains pseudo-labels from detections and camera geometry.Association loss uses the pseudo-label assignment and confidence matrices.Source: A TrackFish3D Training Algorithm · p. 16
  • Associated detections are grouped and triangulated, with low-confidence groups pruned before group embeddings enter contrastive learning.The contrastive objective uses positive pairs within each group.Source: A TrackFish3D Training Algorithm · p. 16
  • The temporal stage predicts embeddings across frames; during warmup, only association loss is active before the full loss is optimized.Parameters are updated with AdamW, and the group representation is appended to memory.Source: A TrackFish3D Training Algorithm · p. 16

B Network Architecture

TrackFish3D combines a geometry-only encoder, global association transformer, and temporal predictor. Its training and inference procedures use calibrated multi-view detections to form 3D groups and link identities across frames.

B Network Architecture

  • The geometric encoder embeds each detection’s location, scale, camera, confidence, and world ray; a global transformer processes detections jointly and predicts cross-camera associations.The model has approximately 655K trainable parameters and no visual backbone.Source: B Network Architecture · pp. 16–18

C Training Details

  • Training runs 400 epochs, with a 100-epoch association-loss warmup before jointly optimizing association, contrastive, and temporal losses.Pseudo-labels from triangulation and bidirectional reprojection checks are cached; sequential frame processing supports sliding-window temporal memory.Source: C Training Details · pp. 18–19

D Inference Implementation

  • Inference groups cross-camera detections, triangulates valid groups, and uses Hungarian matching to link active tracks; the temporal predictor rolls embeddings forward across gaps.Tracks unobserved for more than 15 frames are killed, and short-lived fragments can be merged using distance and embedding similarity.Source: D Inference Implementation · pp. 19–20

E Dataset Statistics

  • Dataset difficulty is characterized by OccScore, the average maximum pairwise bounding-box IoU, and %Overlap, the fraction of frame-camera instances with overlap.The dataset summary also reports fish counts and occlusion statistics.Source: Table 8 · p. 21

F Camera Calibration Parameters

  • Camera calibration is specified by intrinsics and extrinsics; SynFish test scenes include an unseen rig, while 3D-ZeF uses two cameras with sequence-specific extrinsics.SynFish rigs define three cameras and share intrinsics; the 3D-ZeF cameras have distinct intrinsics.Source: F Camera Calibration Parameters · pp. 21–23, Table 9 · p. 21

G Per-Scene Results: SynFish Test Set

  • The SynFish test-set tables report per-scene accuracy and identity-stability metrics for six held-out scenes spanning four to sixteen fish.Scene labels correspond to the test scenes listed in the table caption.Source: Table 12 · p. 23, Table 13 · p. 23

H Per-Scene Results: 3D-ZeF Test Set

  • The 3D-ZeF per-scene table reports tracking results for two five-fish sequences; ReST produces no valid cross-camera associations on zebrafish sequences.Source: Table 14 · p. 24

I Temporal Linking Weight Ablation

Temporal linking combines 3D proximity and embedding agreement during training and inference. Ablations find stable performance across a range of weights, with identity stability degrading when the distance signal is removed.

I Temporal Linking Weight Ablation

  • Training weights α=0.6 and β=0.4 select temporal positive pairs using both spatial proximity and embedding agreement, down-weighting close but dissimilar pairs.This is intended to reduce noisy gradients in temporal prediction.Source: I Temporal Linking Weight Ablation · p. 24
  • Performance is stable for α ∈ [0.4, 1.0], while α=0, β=1 increases identity switches and fragmentation by removing the 3D distance signal.The selected α=0.6, β=0.4 matches training and inference and retains embedding similarity as a complementary cue.Source: I Temporal Linking Weight Ablation · p. 24, Table 15 · p. 24

J Robustness to Camera Calibration Noise

TrackFish3D tolerates small camera calibration errors, but larger extrinsic perturbations reduce tracking performance. The authors attribute the decline to faulty reprojection-based validation and recommend calibration accuracy around σR ≲0.1° and σt ≲0.1 world units.

J Robustness to Camera Calibration Noise

  • The test injects independent Gaussian noise into camera rotations and translations at inference, while keeping the model trained on unperturbed calibration.Four noise levels are compared with the unperturbed baseline on the SynFish test set.Source: J Robustness to Camera Calibration Noise · p. 24, Table 16 · p. 25
  • At σ=0.2, MOTA falls by 14.5 percentage points and MT% drops to 82.3% as reprojection validation rejects more valid associations.At σ=0.05, MOTA drops by 1.7 pp; at σ=0.1, it drops by 3.7 pp with no increase in identity switches.Source: J Robustness to Camera Calibration Noise · p. 25
  • The authors recommend calibration accuracy around σR ≲0.1° and σt ≲0.1 world units for near-baseline performance.They attribute sensitivity to extrinsic errors propagating through DLT triangulation and increasing with object distance.Source: J Robustness to Camera Calibration Noise · p. 25

K Robustness to Detection Noise

The section reports robustness tests and extended implementation details for TrackFish3D. Moderate detection noise is tolerated, while severe noise degrades performance, especially in dense scenes; the extended method describes geometric grouping, contrastive supervision, temporal prediction, and fragment stitching.

K Robustness to Detection Noise

  • The stress test adds false negatives and randomly located false positives to already noisy detector outputs, evaluating four noise levels on six SynFish test scenes.At σ=0.10, MOTA drops by 11.7 pp and identity switches rise to 7.0; at σ=0.20, fragments rise to 24.8 from 6.5.Source: K Robustness to Detection Noise · pp. 25–26
  • At σ=0.40, MOTA drops to 45.4% and MT% to 40.6%, with stronger degradation in the three 16-fish scenes.At σ=0.05, MOTA drops by 5.5 pp; at σ=0.20, it drops by 23.5 pp while no tracks are mostly-lost.Source: Table 17 · p. 25, K Robustness to Detection Noise · pp. 25–26

L Runtime Performance

  • Runtime and memory are reported for a single NVIDIA A100 80GB PCIe GPU with batch size 1, but no measured values are supplied here.Training time covers 400 epochs on SynFish; inference speed is measured per frame across all cameras.Source: L Runtime Performance · p. 26, Table 18 · p. 26

M.1 Grouping, Thresholding, and Geometric Validation

  • Grouping combines learned correspondence probability and transformer-embedding cosine similarity, then matches detections across camera pairs and geometrically validates triangulated groups.The affinity weights are 0.6 and 0.4; groups failing reprojection containment are rejected.Source: §M.1 · p. 27

M.2 Positive Mining for the Contrastive Loss

  • Within-frame detections in the same multi-view group provide contrastive positive pairs without identity labels, while other detections serve as negatives.The denominator includes all other detections, pushing different-identity embeddings apart.Source: §M.2 · pp. 27–28

M.3 Fallback Matching and Temporal Predictor Details

  • A temporal predictor is trained on adjacent-frame matches and fallback matches from a window of up to 10 frames to support linking across transient occlusions.Match weights use the lower group confidence, and predicted embeddings are rolled forward only at inference.Source: §M.3 · p. 28

M.4 Tracker and Fragment Stitching Details

  • At inference, the tracker forms triangulated hypotheses, links them online using predictor-based embedding rollforward, and stitches fragments using learned representations.Stitching adds no learned parameters or extra supervision; optional interpolation and ghost-track removal create no new identity correspondences.Source: §M.4 · pp. 28–29

N.1 Datasets and difficulty characterisation

The evaluation uses synthetic SynFish and real-world 3D-ZeF zebrafish data, with calibrated cameras and held-out test splits. Thirteen end-to-end tracking pipelines are compared under a shared protocol, and TrackFish3D is trained using only calibrated cameras and 2D observations.

N.1 Datasets and difficulty characterisation

  • SynFish contains 14 scenes with 4–16 visually similar fish across three synchronized cameras, with six training scenes held out for testing.Test rigs include one configuration unseen during training.Source: §N.1 · p. 30
  • The real-world 3D-ZeF benchmark uses two calibrated cameras and 2- or 5-fish scenes, with zebra1 and zebra3 for training and zebra2 and zebra4 for testing.SynFish provides ground-truth 3D trajectories for quantitative evaluation; all evaluated datasets provide calibrated cameras.Source: §N.1 · p. 30
  • Difficulty is characterized using an occlusion score based on maximum pairwise box overlap and the percentage of instances with any overlap.The occlusion score averages across frames and views.Source: §N.1 · p. 30

N.2 Baselines and fair 3D tracking pipelines

  • Because no prior method solves this setting end-to-end without identity supervision, the authors compare 13 complete pipelines assembled from strong components.The baselines cover geometric and learned multi-view association, single-view MOT lifted to 3D, and multi-camera tracking followed by triangulation.Source: §N.2 · pp. 30–31
  • The shared evaluation protocol applies the same post-processing to all pipelines, while external components are retrained or fine-tuned on the corresponding training splits.The shared post-processing interpolates short gaps between already-associated observations and removes short-lived ghost tracks.Source: §N.2 · p. 31

N.3 Implementation details

  • TrackFish3D is trained for 400 epochs using calibrated cameras and 2D observations, without identity labels or ground-truth 3D trajectories.All reported results use SAM 3 or YOLO26+SAM2 detections rather than ground-truth boxes.Source: §N.3 · p. 31

O Proof-of-principle downstream analysis on Abnormal Behaviour Detection

The appendix tests downstream behavior analysis, cross-species transfer, kinematic tracking, density scaling, and pseudo-label quality. The results include high window-level anomaly-detection scores from short 3D trajectories, but also identify limits from extreme density and noisy geometric pseudo-labels.

O Proof-of-principle downstream analysis on Abnormal Behaviour Detection

  • 93.9% F1: the z-score detector identifies the sick fish in 92.2% of three-frame windows using TrackFish3D trajectories.The three-frame evaluation yields 359 overlapping windows; Isolation Forest and LOF achieve 78.8% and 80.2% F1, respectively.Source: O Proof-of-principle downstream… · p. 32, Table 20 · p. 33
  • The best single-view consensus F1 is 52.1% versus 80.8% for 3D trajectories, while the best single-view z-score F1 is 85.2% versus 93.9% for 3D.Cam 1 has 0.0% z-score F1, and Cam 2 has 25.9% consensus F1.Source: O Proof-of-principle downstream… · p. 32

P Cross-species generalisation to real-world pigeons

  • On held-out pigeon sequence seq48, TrackFish3D achieves 91.9% MOTA and 94.5% recall, while MOTIP leads on IDF1 and identity-switch count.TrackFish3D reduces fragmentation from 6 to 4 relative to the strongest competing baselines.Source: P Cross-species generalisation to… · p. 34

Q 3D-SORT Kinematic Baseline

  • The 3D-SORT Kalman baseline bridges short gaps under smooth motion but remains 8.1 percentage points below TrackFish3D on SynFish.On 3D-ZeF it lowers MOTA as identity switches increase, particularly when visually identical fish cross or turn unpredictably.Source: Q 3D-SORT Kinematic Baseline · p. 35, Table 22 · p. 35

R Density and Computational Scaling

  • The 30-fish result is a synthetic single-scene stress test, not evidence of a universal capacity threshold or real high-density generalization.At high density, occlusion, missed detections, and plausible cross-view matches can make reliable triangulation impossible when too few fish are visible in two cameras.Source: R Density and Computational Scaling · p. 35, Table 23 · p. 35

S Pseudo-Label Noise Audit

  • Geometry-derived pseudo-labels have declining precision as density increases, despite complete recall in the evaluated scenes.The authors characterize them as noisy, high-coverage supervision rather than clean or intrinsically stable labels.Source: S Pseudo-Label Noise Audit · p. 36, Table 25 · p. 36

T Training-Time Calibration Noise

Training with perturbed camera calibration degrades performance, though tracking remains useful under the tested perturbation. The experiment confirms calibration quality as a practical limitation, not calibration invariance.

T Training-Time Calibration Noise

  • The controlled 3D-ZeF experiment adds Gaussian rotation and translation noise with σR=σt=0.1 to training extrinsics, then evaluates with clean calibration.Geometry-derived supervision is regenerated under the perturbed calibration; Tab. 26 reports performance for this setup.Source: T Training-Time Calibration Noise · p. 36, Table 26 · p. 36
  • Training-time calibration noise degrades performance, but tracking does not collapse and no trajectories become mostly lost.The authors conclude that useful performance remains under the tested perturbation.Source: T Training-Time Calibration Noise · p. 36
  • The experiment does not establish calibration invariance and identifies calibration quality as a practical limitation.Training-time calibration noise causes a clear degradation relative to clean training.Source: T Training-Time Calibration Noise · p. 36

Related papers

Loading 2609.38347v1…