acceptodds
Under review as a conference paper at ICLR 2027

Learning to Track from Privileged Target Appearances

Abstract

Target templates define what a visual tracker searches for, yet the templates available at inference trade off localization certainty with appearance freshness: the initial ground-truth template is exact but becomes stale, whereas recent templates better reflect the current appearance but are cropped from uncertain predictions. We quantify this bottleneck with a non-deployable oracle that supplies an exact current-frame target crop, improving AUC on LaSOT by 15.2 percentage points. This gap reveals a training-only opportunity: frame-level ground truths provide exact current- and future-frame target crops, although such crops are unavailable at deployment. We introduce Privileged Appearance Transfer for Tracking (PATT), a teacher-student training framework that transfers these privileged appearances to a deployable tracker through multi-level representation prediction. The privileged teacher observes exact target crops from past, current, and future frames, whereas the student receives only past-frame templates and learns to predict the teacher's search representations. To avoid transferring unreliable teacher signals, PATT weights this transfer by the teacher’s relative localization advantage over the student and its absolute localization accuracy. After training, the teacher, latent predictor, reliability weights, and privileged crops are removed, leaving standard student-only inference. Across seven benchmarks at two model scales, PATT achieves consistent gains under both long- and short-term tracking protocols. Code and models will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Visual Object Tracking, Learning with Privileged Information

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_xxPUOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies a template acquisition bottleneck in visual tracking. The initial template is accurately localized but becomes outdated, while later templates capture appearance changes but may contain localization errors. The authors use a current-frame oracle unavailable at test time to measure this gap. They then propose PATT, where a teacher uses accurate target crops from past, current, and future frames, while a student uses only historical templates. The student predicts the teacher’s multi-layer search features, with the training loss weighted by the teacher’s relative advantage and absolute IoU. Experiments show consistent gains on short- and long-term benchmarks.

Critical strengths
  1. The paper clearly defines the inference-time template acquisition bottleneck and uses a non-deployable oracle to turn the intuition about template staleness and localization uncertainty into a measurable research question.

  2. The experiments are broad in scope, covering seven benchmarks, two model scales, and multiple evaluation regimes.

Critical weaknesses

Table 2 shows that adding an EMA teacher and a representation prediction branch brings little improvement without accurate target crops from the current and future frames. This suggests that the gains mainly come from privileged appearance information. However, the tracking loss in Eq. (11) directly optimizes only the student, while the teacher is updated through an EMA of the student’s parameters. The paper does not report the teacher’s AUC, how often it outperforms the student, or the fraction of samples with nonzero reliability weights. It therefore remains unclear whether the teacher consistently outperforms the student when given privileged inputs, and how many training samples provide useful auxiliary supervision.

Questions for authors
  1. Please report how the mean IoU difference utea−ustuu_{tea}-u_{stu} and the fraction of samples with nonzero reliability weights change across training epochs.

  2. Historical templates are cropped using ground-truth boxes during training. Are localization perturbations applied to the historical templates during training, beyond the jitter applied to the search crop? Do the teacher and student use the same template-cropping and augmentation procedures?

2 comments
AuthorsOpenReview ↗

Response to Reviewer xxPU: teacher advantage and effective supervision during training, and template cropping

Thank you for the careful review. We agree with your reading of Table 2 that the gain mainly comes from privileged appearance, which is the paper's central claim. Your question is the right test of this claim, and we report the requested statistics below.

W and Q1. Teacher AUC, teacher advantage, and fraction of non-zero weights. With privileged current- and future-frame crops, the final EMA teacher reaches 88.3 LaSOT AUC, a non-deployable setting. With causal (past) templates, the same weights reach 75.4, identical to the student, so the teacher's advantage comes from its privileged inputs. Over training, we record the following statistics on all training samples of PATT-B256. Since utea>ustu≥0u_\mathrm{tea}>u_\mathrm{stu}\ge 0 implies utea>0u_\mathrm{tea}>0, Eq. 10 gives w>0  ⟺  utea>ustuw>0 \iff u_\mathrm{tea}>u_\mathrm{stu}.

Epoch1246810
Mean ustuu_\mathrm{stu}0.7650.8140.8290.8350.8390.843
Mean uteau_\mathrm{tea}0.7480.8630.8810.8880.8920.894
Mean utea−ustuu_\mathrm{tea}-u_\mathrm{stu}−0.0170.0490.0520.0530.0530.051
Fraction with w>0w>054.2%75.4%79.7%81.2%81.9%81.0%
Mean ww0.2310.3060.3180.3230.3250.313
  1. After the first epoch, the teacher is better. From epoch 2 on, the privileged teacher's mean IoU exceeds the student's by 0.049–0.053, and it is more accurate on 75–82% of samples. In epoch 1, the teacher lags slightly (mean gap −0.017), because its weights are a moving average of the student's and the student improves fastest early in training. The reliability weight is designed for such cases. It suppresses transfer on samples where the teacher is not better.
  2. The advantage persists as the student improves. The student's mean IoU rises from 0.765 to 0.843 over training, and the teacher stays about 0.05 ahead of it from epoch 2 on. The auxiliary supervision therefore remains informative for the rest of training (mean ww 0.31–0.33 from epoch 2 on).
  3. The weighting is selective. After epoch 1, 18–25% of samples receive zero weight because the teacher is no better than the student on them. Uniform weighting would still push the student toward these targets, which is consistent with it being worse (Table 2, row 14 of the revised PDF). Our weighting instead focuses transfer on the samples where the teacher is better.

We have added these statistics to Appendix G of the revised PDF.

Q2. Template perturbation and cropping procedures.

  • Historical templates. We crop them from ground-truth boxes without box jitter, a simplification that keeps the framework clean and avoids extra jitter parameters to tune. We apply translation and scale jitter only to the search crop.
  • Teacher and student use the same procedure. The first K−2K-2 teacher templates are the same tensors as the student's, and the privileged crops use the same crop operator (factor 2, 128×128) and the same augmentation. Both branches share the same search crop. The only difference is which frames fill the last two slots, which is what makes the past-only teacher a clean control. We have made this explicit in Appendix A of the revised PDF.

We hope these statistics resolve your concern, and we would appreciate your considering them in the final assessment.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer xxPU,

Thank you again for your careful review. As the discussion period ends soon, we would like to check whether our response addresses your questions, in particular the teacher statistics across training (Appendix G) and the template-cropping procedure (Appendix A). We would be glad to clarify anything further.

Thank you very much.

Reviewer_iMkYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Privileged Appearance Transfer for Tracking (PATT), a teacher–student training framework for single-object tracking. The method tackles the difficulty of obtaining accurate and up-to-date target templates at inference time by leveraging ground-truth target appearances that are available only during training. Specifically, the privileged teacher is given target crops from the current and future frames, whereas the student has access only to historical templates. A Transformer-based predictor aligns the student's multi-level search representations with those of the teacher. In addition, a reliability-aware weighting scheme modulates the distillation loss according to the teacher's relative localization advantage and absolute localization accuracy. The framework adds no inference-time computation. Experiments on seven tracking benchmarks and two model scales show consistent gains, and ablation studies examine privileged template selection, representation prediction, and reliability weighting.

Critical strengths

1.PATT uses additional target appearance information during training without increasing inference cost. It delivers meaningful gains on LaSOT and LaSOText and remains competitive across benchmarks and model scales. 2. A past-only teacher offers limited gains, while current-frame privileged target appearances substantially improve performance. Ablations on output imitation, feature matching, multi-level prediction, and reliability weighting further support their contribution.

Critical weaknesses
  1. PPKD [1] transfers future privileged information through intermediate representation distillation with a learnable feature mapping. UETrack [2] adapts teacher supervision using ground-truth-based reliability estimates, conceptually similar to PATT’s localization-quality weighting. Although PATT introduces privileged target crops and target-conditioned search representations, its methodological contribution beyond combining these strategies remains unclear.
  2. The ablations show improvements over several distillation baselines but do not establish the need for a Transformer-based predictor. Comparisons with lightweight feature mappings [1] and teacher-free alternatives, such as intermediate localization or foreground-aware supervision, would clarify whether the gains stem from privileged appearance transfer or additional feature-level supervision.
  3. Including additional recent trackers under comparable training and evaluation settings would help better contextualize PATT’s performance and further strengthen the empirical evaluation.

[1] Peisen Zhao, Lingxi Xie, Jiajie Wang, Ya Zhang, and Qi Tian. “Progressive privileged knowledge distillation for online action detection.” Pattern Recognition, vol. 129, article 108741, 2022. [2] Ben Kang, Jie Zhao, Xin Chen, Wanting Geng, Bin Zhang, Lu Zhang, Dong Wang, and Huchuan Lu. “UETrack: A Unified and Efficient Framework for Single Object Tracking.” In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026.

Questions for authors
  1. Could the authors clarify the tracking-specific methodological insight that distinguishes PATT from PPKD beyond ground-truth target crops and a Transformer predictor?
  2. Could the authors compare PATT with auxiliary localization or foreground-aware feature supervision under matched training conditions?
  3. Could the authors clarify how appearance information and localization-related cues from ground-truth target crops contribute separately to the oracle performance gap?
4 comments
AuthorsOpenReview ↗

Response to Reviewer iMkY (1/3): tracking-specific insight vs. PPKD and UETrack

Thank you for the detailed review and the references. We address each point below.

W1 and Q1. What is new. Learning with privileged information, distillation, and per-sample weighting of teacher signals are general tools, and PPKD and UETrack are good instances of them. PATT does not contribute a new tool. It gives a tracking-specific answer to two questions that these tools leave open: how to build privileged information, and how to transfer it.

  • How to build it. A tracker finds the target by its appearance, given by its templates. How the target looks now is therefore highly informative. Deployment cannot provide it, since cropping the current target needs the very box the tracker is predicting. During training, box annotations provide this crop for every frame. PATT gives exact current- and future-frame target crops to a teacher with the same network and template budget, so the teacher is the student itself under perfect template acquisition. Our oracle shows how informative such a crop is (+15.2 AUC on LaSOT).
  • How to transfer it. The privileged crops exist only in training, so the student must learn their effect without ever seeing them. PATT trains the student, which sees only past templates, to predict the privileged teacher's search representations at multiple encoder layers, with a per-sample reliability weight. Table 2 of the main paper measures what each choice contributes. Predicting representations gives 75.4 / 55.4, compared with 73.5 / 53.7 for imitating the teacher's box output and 73.3 / 54.3 for forcing the student's features to equal the teacher's.

The same table shows that the gain comes from the privileged crops rather than from the general tools. The past-only teacher uses all other parts of PATT (the same predictor, loss and weighting) and gives +0.2 / −0.1, while adding the privileged crops gives +2.4 / +2.7.

Differences from PPKD and UETrack.

PPKDUETrack (TAD)PATT
Taskonline action detection (per-frame action classification)efficient multi-modal trackingsingle-object tracking
Teacherfour offline teachers, trained beforehand, used in a four-stage curriculumSUTrack-B, pretrained and frozenEMA of the student with privileged inputs, no separate teacher training, single stage
Teacher's advantagemore of the video (future frames of the whole scene)more capacity (a larger tracker on the same inputs)a better view of the target (exact current- and future-frame target crops)
Per-sample controlnonelearned binary gate (Adaptive Net, Gumbel-Softmax, surrogate loss)continuous weight from both branches' IoU (advantage × quality), no learned parameters
  1. A better view of the target, not more of the video. PPKD's privileged information is more of the video: future frames of the whole scene. PATT's privileged information is a better view of the target: exact crops of the target in the current and a future frame. This matches the nature of tracking, which follows a single target through a video and identifies it by its appearance. The past+current teacher (Table 2, row 4 of the main paper) separates the two. It adds only the exact target crop of the current frame, which the student already sees, and still obtains +2.1 / +2.3 of the full +2.4 / +2.7 gain. The gain therefore comes from the better view rather than from more of the video.
  2. Reliability weighting measures the value of the privileged crops. UETrack has a similar motivation, suppressing unreliable teacher signals. The weighting differs in three ways. (a) Teacher and student share weights (up to EMA), search crop and template budget, so utea−ustuu_\mathrm{tea}-u_\mathrm{stu} measures how much the privileged crops helped on each sample. Our weight thus asks whether the privileged crops carried useful information on this sample, whereas UETrack's gate asks whether a different, larger teacher can be trusted. (b) The weight is continuous. It grows with the teacher's IoU advantage and with its absolute IoU, so a marginal advantage is not treated like a large one, as it would be under UETrack's binary gate. (c) It is computed in closed form from the IoU of both branches, with no extra network, surrogate objective, or separate update.

We have added PPKD and UETrack to Related Work (Section 2.2) and discussed UETrack's gating in Section 3.3 of the revised PDF. Parts 2–3 address your remaining questions.

AuthorsOpenReview ↗

Response to Reviewer iMkY (2/3): privileged appearance vs. extra feature-level supervision, and predictor form

W2 and Q2. Matched teacher-free controls and predictor forms. All rows use the PATT-B256 setting with the same data, schedule, initialization, four selected encoder layers, and deployed tracker. LaSOT / LaSOT$_\mathrm{ext}$ AUC:

VariantPrivileged cropsLaSOTLaSOT$_\mathrm{ext}$
Student-only baseline–73.052.7
+ intermediate localization supervision (center head on each selected layer)–72.852.4
+ foreground-aware token supervision (per-token foreground/background BCE on each selected layer)–72.552.0
+ past-only EMA teacher, same 12-layer predictor, loss and weighting–73.252.6
Privileged teacher, per-token linear map (PPKD-style)✓73.354.5
Privileged teacher, 2-layer Transformer predictor✓74.054.6
PATT (12-layer Transformer predictor)✓75.455.4

Two conclusions follow.

  • Extra supervision does not explain the gain. Intermediate localization and foreground supervision at the same layers even lower AUC slightly (−0.2 / −0.3 and −0.5 / −0.7). The past-only teacher keeps all of PATT's feature-level supervision, including the same predictor, and drops only the privileged crops. It gives +0.2 / −0.1.
  • The Transformer predictor is needed for the full gain, but only with privileged crops. With privileged crops, the gain grows with predictor capacity: +0.3 / +1.8 for a per-token linear map, +1.0 / +1.9 for a 2-layer Transformer, and +2.4 / +2.7 for PATT's 12-layer Transformer. Without privileged crops, the same 12-layer predictor gives +0.2 / −0.1. The predictor is discarded after training, so it adds no inference cost.

We have added these rows to Table 2 of the revised PDF (rows 6–7 and 12–13).

AuthorsOpenReview ↗

Response to Reviewer iMkY (3/3): decomposing the oracle gap, and recent trackers

Q3. Appearance vs. localization cues in the oracle gap. We use the same student-only network (73.0 on LaSOT) and change only how the dynamic template is obtained, along two axes: freshness (which frame it shows) and localization certainty (which box crops it). All rows except the first are non-deployable diagnostics.

Template suppliedFreshExact boxLaSOT AUCGain
Deployed protocol––73.0–
Dynamic templates re-cropped with GT boxes (stale)–✓80.1+7.1
Current frame, cropped with the tracker's own predicted box✓–71.8−1.2
Current frame, GT center but predicted size/aspect✓center only78.9+5.9
GT crop from frame t−1t-1 (no pixels shared with the search frame)✓✓86.1+13.1
Oracle (Figure 1)✓✓88.2+15.2
  • Fresh appearance and exact cropping both matter. An exact but stale template recovers +7.1, while a fresh template cropped with the tracker's own box even loses 1.2, because its localization errors enter the template. On the current frame, the gain grows with crop exactness: −1.2 with the predicted box, +5.9 with the GT center, and +15.2 with the GT box. This matches the bottleneck we define: at deployment, no template is both fresh and localization-certain.
  • Most of the gap remains without pixel overlap with the search frame. A GT crop from the previous frame t−1t-1 shares no pixels with the search frame, yet it recovers 86% of the gap (+13.1 of +15.2).

We have added this analysis to Appendix D of the revised PDF.

W3. More recent trackers. We compare with four recent trackers, MCITrack [1], SPMTrack [2], DTPTrack [3] and TGTrack [4], using the AUC reported in their papers (search resolution in the name).

TrackerLaSOTLaSOT$_\mathrm{ext}$TrackingNet
MCITrack-B22475.354.686.3
SPMTrack-B37874.9–86.1
DTPTrack-B22474.3–85.6
TGTrack-B25674.653.286.2
PATT-B25675.455.486.6
MCITrack-L38476.655.787.9
SPMTrack-L37876.8–86.9
DTPTrack-L37877.5–86.9
TGTrack-L25675.854.887.5
PATT-L25676.756.787.4

At base scale, PATT-B256 has the best AUC on all three benchmarks. At large scale, PATT-L256 has the best LaSOT$_\mathrm{ext}$ AUC among these trackers, while trackers with larger 378/384 inputs lead on LaSOT (DTPTrack-L378, SPMTrack-L378) and TrackingNet (MCITrack-L384).

When the input resolution is matched, PATT-L256 exceeds TGTrack-L256 by +0.9 / +1.9 on LaSOT / LaSOT$_\mathrm{ext}$ and is on par on TrackingNet (−0.1). We have added these trackers to Table 1 of the revised PDF.

We hope these results address your questions on the source of the gain, the oracle gap, and recent trackers. If so, we would appreciate your considering an updated assessment, and we are happy to answer further questions.

References

[1] Ben Kang, Xin Chen, Simiao Lai, Yang Liu, Yi Liu, and Dong Wang. Exploring enhanced contextual information for video-level object tracking. In AAAI, 2025.

[2] Wenrui Cai, Qingjie Liu, and Yunhong Wang. SPMTrack: Spatio-temporal parameter-efficient fine-tuning with mixture of experts for scalable visual tracking. In CVPR, 2025.

[3] Yuqing Huang, Liting Lin, Weijun Zhuang, Zhenyu He, and Xin Li. Drift-resilient temporal priors for visual tracking. In CVPR, 2026.

[4] Wanting Geng, Xin Chen, Chuanyu Sun, Jie Zhao, Ben Kang, Dong Wang, and Huchuan Lu. TGTrack: Temporal generative learning for unified single object tracking. In CVPR, 2026.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer iMkY,

Thank you again for your review. As the discussion period ends soon, we would like to check whether our response addresses your concerns, in particular the matched teacher-free controls and predictor forms (Table 2, rows 6–7 and 12–13), the comparison with PPKD and UETrack (Sections 2.2 and 3.3), the oracle decomposition (Appendix D), and the recent trackers added to Table 1. We would be glad to clarify anything further.

Thank you very much.

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

The paper shows that trackers are limited by stale or noisy templates (an oracle with the GT current-frame crop gives +15.2 AUC on LaSOT) and proposes PATT, a training-only teacher-student scheme. An EMA teacher sees GT crops from the current and a future frame, and the student learns to predict its multi-level search features through a latent predictor, with a reliability weight based on IoU. Nothing changes at inference. Results on seven benchmarks at two scales, with ablations.

Critical strengths

Simple idea with no inference cost, applicable to any template-based tracker. The past-only teacher ablation is a good control showing the gain comes from the privileged crops and not from EMA self-distillation.

Critical weaknesses
  • There is no sensitivity analysis for the loss hyperparameters lambda, tau, gamma and the EMA momentum rho. Since several margins over the strongest competitor are small (e.g. +0.1 and +0.2 AUC on TrackingNet), it is hard to tell how robust the gains are to these choices.
  • Results come from single runs on a single base tracker. Reporting variance over a few seeds, at least for the B model on LaSOT, would make the comparison with RELO more convincing.
Questions for authors
  • How sensitive are the results to lambda, tau, gamma and rho? Were these tuned on LaSOT test?
  • How much does the gain depend on the temporal offset between the current and the future frame?
  • What is the student-only baseline for PATT-L256, so that the gain at large scale can be measured against its own baseline?
14 comments
Area_Chair_5daPOpenReview ↗

a question regarding the paper

None of the reviewers raised this issue, so I wanted to address it directly. The paper states that by using and exact oreacle, AUC score on LaSOT increases from 0.73 to 0.882 at both abstract and introduction; however, Table 1 reports a value of 0.767. This discrepancy is quite confusing and needs clarification. Does this mean that the proposed method is not good to introduce good templates during inference? The reported results is very close to the existing sota trackers.

AuthorsOpenReview ↗

Response to Area Chair 5daP: oracle numbers vs. Table 1

Thank you for raising this. The two numbers come from different settings.

  • 73.0 → 88.2 is a non-deployable oracle, not PATT's result. Both numbers come from the same student-only B256 tracker, which has the architecture of PATT-B256 but is trained without PATT. 73.0 is its standard inference result. 88.2 is obtained when, at every frame, we supply it with an exact template cropped from that frame with its ground-truth box. This crop requires the very box the tracker must predict, so it is unavailable at deployment. The oracle only measures how much an exact, up-to-date template is worth, and it motivates PATT.
  • Table 1 reports deployable trackers. PATT uses such crops only during training. At inference, it runs the standard tracker with no ground truth beyond the first frame. PATT-B256 reaches 75.4 AUC on LaSOT, +2.4 over the same 73.0 baseline, and PATT-L256 reaches 76.7.

We hope this clarifies the difference. Thank you again.

Area_Chair_5daPOpenReview ↗

reply for the answer

I know that the first result represents an idealized upper bound obtained using a perfect oracle. However, given the performance achieved by the perfect oracle and the substantial gap between that result and your method, it appears that the student network performs considerably worse than the oracle. I am also concerned that the reported improvements may be largely attributable to the backbone architecture rather than the proposed method itself. You use Fast-iTPN encoders, which are relatively uncommon backbones in this field. Interestingly, ReLO also uses the same backbones and achieves quite competitive results, which further raises the question of whether the improvements are primarily due to the backbone rather than the proposed approach. Could you clarify whether you also evaluated your proposed method using DINOv2-Base and DINOv2-Large? Several of the competing methods reported in your comparison table use these backbones. Evaluating your method with the same backbones would provide a fairer comparison and help isolate the actual contribution of your proposed approach.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers and the Area Chair for their constructive feedback. We are encouraged that the reviewers found the problem "clearly defined" and made "a measurable research question" by the oracle (xxPU), the method a "simple idea with no inference cost" (oDoD), the past-only teacher "a good control showing the gain comes from the privileged crops and not from EMA self-distillation" (oDoD, and similarly iMkY), and the experiments "broad in scope" (xxPU).

Core idea. A tracker finds the target by its appearance, given by its templates, and the most useful template shows how the target looks now. A deployed tracker can never reliably obtain such a crop, since it would need the very box it is trying to predict. During training, box annotations provide it for every frame. PATT gives exact current- and future-frame target crops to a teacher with the same network (EMA weights) and the same four template slots, so the teacher is the student under perfect template acquisition. The student learns to predict the search representations this teacher forms. After training only the student is kept, with unchanged architecture, inputs and speed.

Two shared questions. The reviews ask (i) whether the gain comes from privileged appearance rather than from extra supervision, the predictor, hyperparameter choice, or seed variance (iMkY, oDoD), and (ii) what the privileged inputs actually provide, to the teacher and in the oracle (xxPU, iMkY). We tested each point. Numbers below are LaSOT / LaSOT$_\mathrm{ext}$ AUC for PATT-B256 unless noted, with details under each review.

(i) The gain comes from privileged appearance.

  1. Not extra supervision or the predictor (iMkY). As oDoD and iMkY noted, the past-only teacher in Table 2 is a first control. It keeps PATT's EMA teacher, 12-layer predictor, loss and weighting, and drops only the privileged crops. It gives +0.2 / −0.1, whereas PATT gives +2.4 / +2.7. Teacher-free intermediate localization and foreground supervision at the same layers do not help either (−0.2 / −0.3 and −0.5 / −0.7). The predictor matters only together with the privileged crops: with them, the gain grows from +0.3 / +1.8 for a per-token linear map to +2.4 / +2.7 for PATT's Transformer predictor.
  2. Robust to hyperparameters and seeds (oDoD). Over 3 seeds, PATT-B256 obtains 75.4 ± 0.1 vs. 72.9 ± 0.2 LaSOT AUC for the baseline, and is better in every paired seed. Across 12 settings of λ,τ,γ,ρ\lambda,\tau,\gamma,\rho, LaSOT AUC stays within 74.9–75.5. The L256 student-only baseline obtains 73.7 / 54.1, so PATT adds +3.0 / +2.6 at large scale. PATT improves its own baseline on all seven benchmarks at both scales.

(ii) What privileged appearance provides.

  1. A teacher that is better than the student (xxPU). After the first epoch, the privileged teacher localizes better than the student on 75–82% of training samples, with a stable mean IoU gap of about 0.05.
  2. An oracle gap that needs both fresh appearance and exact cropping (iMkY). In the oracle study, an exact but stale template recovers +7.1 of the 15.2 points, and a current-frame template cropped with the tracker's own predicted box even loses 1.2.

Revision. We have uploaded a revised PDF, with changes marked in blue. Sections 2.2 and 3.3 relate PATT to PPKD and UETrack, Table 2 adds the teacher-free controls, the future-only teacher and the predictor forms, and Table 1 adds more trackers. Appendices C–G add results with a DINOv2-B backbone, the oracle decomposition, the robustness analyses, the student-only baselines at both scales, and the teacher statistics.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.