acceptodds
Withdrawn from ICLR 2027

DASTrack: Exploiting DINOv3 Features, Adaptive Modulation, and Spatial Refinement for Visual Tracking

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Visual object tracking has benefited from adopting pre-trained Vision Transformers as backbones. While the recently introduced DINOv3 offers rich intermediate-layer representations, it has not been specifically adapted for tracking, and existing backbone adaptation strategies are insufficient to fully exploit its potential. Moreover, existing feature modulation strategies rely on template-search interaction and coordinate-based trajectory encoding, lacking explicit channel-wise conditioning on target appearance and trajectory modulation that accounts for spatial response distributions. To address these limitations, we propose DASTrack, a framework that exploits DINOv3 features for tracking through dedicated backbone adaptation, adaptive modulation, and spatial refinement. For backbone adaptation, we propose dense LoRA, which jointly encodes template and search images and aggregates multi-layer features, exploiting reliable intermediate representations of the backbone. To explicitly condition features on the target, we introduce template-adaptive modulation and trajectory-guided modulation to enhance search features based on target appearance and historical response distributions, respectively. We further design a spatial refinement head that iteratively refines search features through Gaussian-gated attention to achieve precise localization. We evaluate DASTrack on seven benchmarks, achieving new state-of-the-art results of 81.4% AO on GOT-10k and 72.6% AUC on UAV123, with competitive results on other datasets. Code and results will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Visual object tracking, DINOv3, Dense LoRA, Adaptive modulation, Spatial refinement

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_9ycyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript proposes DASTrack, a single‑object tracker that introduces the pre‑trained DINOv3 vision Transformer into single‑object tracking. The authors present Dense LoRA (DLoRA) as the backbone adaptation strategy, which spatially concatenates template and search images and aggregates multi‑layer intermediate features from the backbone. Built upon DLoRA, the paper develops template‑adaptive modulation (TAM), trajectory‑guided modulation (TGM), and a spatial refinement head (SRH) for iterative localization. The model is evaluated on seven tracking benchmarks with promising results, accompanied by abundant experimental analyses.

Critical strengths

This work specifically adapts DINOv3 for single‑object tracking. The proposed Dense LoRA (DLoRA) provides a practical backbone adaptation recipe: spatially concatenating template‑search inputs and aggregating multi‑layer intermediate features of the network. This design aligns with the intrinsic property of DINOv3. This exploration offers valuable references for future tracking research built upon DINO‑family foundation models.

Critical weaknesses

1.The validation of module effectiveness heavily relies on the paper’s multi‑stage training strategy, where previously trained modules are frozen when training new components. There is no end‑to‑end joint‑training baseline. It is therefore difficult to distinguish whether the observed performance gains come from the intrinsic capability of each module, or from residual compensation of later modules for features produced by frozen earlier components.  2.The discussion in the related‑work section is insufficient. The substantial novelty of this paper concentrates on the Dense LoRA (DLoRA) backbone adaptation for DINOv3, while TAM, TGM and SRH are incremental engineering improvements built upon well‑established paradigms. However, the manuscript fails to clearly draw the boundary between these customized modules and existing prior approaches.

Questions for authors

1.Some existing works adopt gaussian response maps. Why does the TGM module not directly use them as the Value in attention? 2.In Table 3, the multi-layer feature fusion (MLF) module of DLoRA shows counterintuitive results and the given explanation is flawed. Since the GOT-10k training set enjoys rich category diversity, intermediate-layer features are expected to be beneficial. Why does multi-layer fusion hurt performance?

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

This paper presents DASTrack, a single-object tracking framework that adapts DINOv3 through Dense LoRA and combines template-adaptive modulation, trajectory-guided modulation, and iterative spatial refinement. The method uses target appearance and historical response distributions to improve feature discrimination and localization. Experiments on seven benchmarks show competitive performance, including 81.4% AO on GOT-10k and 72.6% AUC on UAV123, the best results among the methods reported in the paper. Ablation studies demonstrate improvements over the DINOv3–LoRA baseline, although the benefits of multi-layer fusion and some aspects of the evaluation protocol require clarification.

Critical strengths
  1. On GOT-10k, DASTrack achieves 81.4% AO, exceeding the strongest compared results in Table 1 by 1.4 percentage points. Table 3 further shows that adding TAM, SRH, and TGM improves the DINOv3–LoRA baseline from 79.4% to 81.4% AO. This is important because it demonstrates that the gains are not solely attributable to adopting a stronger pretrained backbone.
  2. TGM retains spatial information about competing target candidates instead of relying only on bounding-box coordinates or a scalar confidence score. Table 5 shows a 0.5-percentage-point AO improvement when full response maps replace scalar classification scores. This provides direct empirical support for a practically relevant design choice.
  3. The paper evaluates seven datasets and investigates component contributions, LoRA rank, fusion layers, refinement iterations, Gaussian spread, and history length. Attribute-level analysis and qualitative examples complement the aggregate results. The discussion also acknowledges limitations under fast motion and the inconsistent benefit of multi-layer fusion, helping establish the method’s current scope.
Critical weaknesses
  1. The methodological novelty is primarily in adaptation and integration, and the necessity of some choices is not fully established.DLoRA combines spatial concatenation, standard LoRA, and multi-layer aggregation, while TAM adopts a FiLM-style affine transformation. This does not diminish the practical value of the framework, but the paper would benefit from clearer positioning of its specific advances and controlled comparisons against simpler alternatives. For example, a comparison between spatial concatenation and separate positional/type embeddings under the same backbone would strengthen the adaptation argument..
  2. The evidence for multi-layer fusion is mixed. In Table 3, adding multi-layer fusion reduces GOT-10k AO from 81.4% to 81.1%, while improving TNL2K AUC by 0.3 percentage points. In Table 7(b), the default fusion configuration and the repeated-final-layer configuration both achieve 81.0% AO. These results suggest that cross-layer aggregation is not consistently beneficial. The paper should qualify this contribution and clarify the conditions under which intermediate-layer features provide an advantage.
  3. Benchmark-specific configurations and some numerical descriptions need clarification. Section 4.1 states that fusion layers and the trajectory Gaussian parameter are adjusted for each benchmark, but the full settings and their selection procedure are not provided. The headline GOT-10k result of 81.4% also differs from the 81.1% result of the full configuration in Table 3. In addition, the conclusion attributes a 1.5-percentage-point TNL2K gain to TAM, TGM, and SRH, whereas Table 3 indicates that these components contribute 1.2 points, with the remaining 0.3 coming from multi-layer fusion.
Questions for authors

Please see weakness above.

Reviewer_RZp4OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose DASTrack which is a single-object tracker built on a frozen DINOv3-L backbone with LoRA adapters (rank 64). It makes four additions: (1) Dense LoRA (DLoRA). The template and search crops are pasted side by side into one image, so DINOv3’s native RoPE tells them apart and no token-type embeddings are needed. LoRA is added to every linear layer. Search tokens from layers 19, 21 and 23 are combined with learned softmax weights. (2) TAM (template-adaptive modulation). This is FiLM: the template tokens are average-pooled, and two linear layers turn the result into a per-channel scale and shift for the search features. (3) TGM (trajectory-guided modulation). The last 8 boxes are drawn as Gaussian maps and stacked with the last 8 response maps. An MLP turns these into “values” for a cross-attention whose queries and keys come from the search features. The result is added back through a residual scaled by a learnable scalar. (4) SRH (spatial refinement head). A center head predicts a box, the box becomes an isotropic Gaussian, and that Gaussian is used as the value in self-attention over the search features. This repeats for M = 2 iterations. Training is conducted in four stages, with the modules trained in each stage subsequently frozen. The method achieves 81.4% AO on GOT-10k and 72.6% AUC on UAV123. While these results are competitive, performance on LaSOT, LaSOT_ext, TrackingNet, TNL2K, and NFS remains below the state of the art.

Critical strengths

Main strengths of the paper can be summarized as follows:

  1. The input design is elegant. Earlier LoRA-based trackers fed the template and search region in as two separate images and had to add extra “token type” embeddings to tell them apart. DASTrack instead pastes the two crops side by side into a single image. DINOv3’s own positional encoding then distinguishes them naturally, and the backbone stays closer to how it was pre-trained. As a side benefit, this uses fewer tokens and makes the model cheaper.
  2. The GOT-10k result is genuinely strong. That benchmark requires training only on its own data and tests on object classes never seen in training, so a high score there says something real about generalization.
Critical weaknesses

Main weaknesses can be summarized as follows:

  1. The most immediate problem is that two of the paper’s headline claims don’t hold up. The authors say they are the first to adapt DINOv3 for tracking, but a paper titled “DINOTrack: Leverage Differential Attention for Noise-Aware Visual Tracking with DINOv3” was published in mid-2026, before this submission. They also claim a new state of the art on UAV123, but their 72.6 beats LoRAT-L by only 0.1, and the larger LoRAT-g-378 reports 0.739 AUC on UAV123.
  2. Beyond using Dinov3 backbone, the method’s novelty is limited. Each module is a known technique moved onto a new backbone: (i) The template modulation is standard FiLM (Feature-wise Linear Modulation) . (ii) Encoding past target positions as Gaussian maps closely resembles ToMP. (iii) Learned fusion of multiple layers was already done in SiamRPN++. (iv) “Dense LoRA” is mostly LoRA on all linear layers, as LoRAT already did, plus layer fusion. Much of the improvement over prior work seems to come simply from switching to DINOv3. The baseline alone already scores close to previous state-of-the-art methods.
  3. The evaluation protocol also raises concerns. The authors state that some hyperparameters are adjusted separately for each benchmark at inference time, and their sensitivity sweeps were run directly on test sets. The GOT-10k headline number appears to come from a configuration without multi-layer fusion, while other benchmarks use the version with it. In effect, the reported numbers mix different models, tuned on the data they’re evaluated on. A fair evaluation would use one fixed configuration chosen on validation data.
  4. Many of the reported gains are also too small to trust without more evidence. Individual modules add between 0.1 and 0.7 points, and different layer choices for fusion give nearly identical results. Without multiple training runs or any measure of variance, it’s hard to tell real improvements from noise.
  5. The training procedure is unusually complicated, with four stages that each freeze everything trained before. The ordering is odd: multi-layer fusion is added last, after the downstream modules have already learned to work with final-layer features.
  6. Several recent competitors are missing from the comparison. These include LoRATv2 and DTPTrack, the latter of which reports a 77.5% Success rate on LaSOT and an 80.3% AO on GOT-10k and outperforms DASTrack on TNL2K. SAM-based trackers aren’t discussed at all, even though they work under a different protocol and have recently posted very high GOT-10k numbers.
  7. Finally, the paper doesn’t really show that its modules work for the reasons it claims. The trajectory module is said to suppress distractors, but nothing in its design is explicitly distractor-aware.
Questions for authors

Please provide more details about the third weakness mentioned above.

2 comments
Reviewer_RZp4OpenReview ↗

my last evaluation

Thank you for the authors' response. Unfortunately, most of my concerns remain unresolved. Therefore, I will maintain my original score.

About the comparison with recent trackers

We thank the reviewer for pointing out ARPTrack-L384 (CVPR 2025). The revised paper now makes no state-of-the-art claims, and we have added ARPTrack-L384 to Table 1 for completeness. ARPTrack-L384 is 0.1% and 0.2% higher in GOT-10k AO and TrackingNet AUC, while DASTrack is 1.3% and 1.8% higher in LaSOT and LaSOT_ext AUC. Our contribution lies not in outperforming every existing tracker, but in exploiting DINOv3 for tracking through our backbone adaptation and tracking-specific modules. On the same DINOv3 backbone, our modules improve GOT-10k AO from 79.4% to 81.4% (Table 3).

About the variance across training runs

The gains of our modules are consistent across independently trained models. In Table 3, the GOT-10k and TNL2K results come from two separately trained model series (GOT-10k-only and full-data training), and TAM, SRH, and TGM each improve both. To support reproducibility, our code and results will be made publicly available.

We thank the reviewer again for the time and effort spent on our submission.

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

Related papers

Discussion (0)

Sign in to comment.