DisDop: Distillation with Domain Priors

• DisDop is a two-stage framework that distills visual, textual, and contextual priors from RemoteCLIP and DINOv3 to enable open-vocabulary aerial object detection.
• The method first performs visual prior distillation on unlabeled data and then integrates textual/contextual supervision during full detector training on labeled images.
• Empirical results demonstrate that DisDop improves both open-set and closed-set detection metrics on benchmarks like DIOR, DOTAv2.0, and LAE-80C compared to previous baselines.

DisDop is a two-stage framework for open-vocabulary aerial object detection (OVAOD) that distills multi-level domain priors from remote sensing foundation models into a lightweight detector. It is introduced as “Distillation with Domain Priors for Open-Vocabulary Aerial Object Detection” and targets a setting in which detectors must recognize aerial objects from natural-language category descriptions, including categories unseen during training. Its central premise is that open-vocabulary detectors designed for natural images transfer poorly to aerial imagery because of domain shifts in scale, orientation, object density, and local appearance; DisDop addresses this by transferring visual, textual, and contextual priors from RemoteCLIP and DINOv3 into an LAE-based detector (Xu et al., 23 May 2026).

1. Problem formulation and design rationale

DisDop is situated in the OVAOD setting, where the detector is not restricted to a predefined closed label set and must instead use text descriptions to detect both seen and unseen aerial categories. The motivating difficulty is that aerial imagery differs substantially from natural-image corpora: the paper highlights a broader range of variations in scales, orientation differences, lack of distinctive appearance features, and the scarcity of large-scale labeled drone-view data. This creates a domain gap for open-vocabulary detectors pretrained on natural scenes (Xu et al., 23 May 2026).

The framework is motivated by a second mismatch as well: even strong vision-LLMs and detection systems often inherit priors from natural-image data, whereas aerial imagery benefits from priors learned directly in remote sensing regimes. DisDop therefore uses remote sensing foundation models rather than relying only on models trained on natural images. In the paper’s formulation, RemoteCLIP contributes cross-modal alignment and text-space semantics, while DINOv3 contributes fine-grained local and spatial feature structure. This suggests a division of labor in which language-aligned semantics and local patch discrimination are distilled separately but coordinated within a single detector (Xu et al., 23 May 2026).

The resulting notion of “multi-level domain priors” has three components. Visual priors come from the RemoteCLIP vision encoder and DINOv3 vision encoder; textual priors come from the RemoteCLIP text encoder and capture inter-category semantic relations; and contextual priors inject global image context into local instance supervision, particularly for small objects. The framework is built on the LAE detector, with a Swin Transformer-Tiny student backbone, and is designed so that the heavy teacher models are needed during distillation rather than inference (Xu et al., 23 May 2026).

2. Overall architecture and two-stage training procedure

DisDop is trained in two sequential stages. Stage I performs Visual Prior Distillation (VPD) on unlabeled LAE-1M images. In this stage, the student backbone is trained to match a fused teacher representation derived from RemoteCLIP-ViT-L/14 and DINOv3-ViT-L/16. Only the student backbone is updated; the teacher side is frozen. The goal is to inject domain-specific visual priors into the backbone before full detector training (Xu et al., 23 May 2026).

Stage II trains the full detector on labeled LAE-1M. The paper states that the objective combines the standard LAE detection losses with a textual/contextual distillation loss, denoted conceptually as

$$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{det}} + \mathcal{L}_{\text{distill}}.$$

The exact decomposition of $\mathcal{L}_{\text{det}}$ and the weighting between detection and distillation losses are not provided in the supplied text, and the paper excerpt likewise does not specify the exact Stage II coefficient schedule (Xu et al., 23 May 2026).

The architecture distinguishes where distillation occurs. Visual distillation acts at the backbone level. Textual prior distillation acts on the detector’s output category features $f_c^i$. Contextual prior distillation modifies the visual supervision target $f_{eh}^j$ used in Stage II. The paper does not state that distillation is performed explicitly at the proposal level, neck level, or box-regression head, so the framework is best described as backbone-level visual distillation plus representation-level semantic and contextual alignment (Xu et al., 23 May 2026).

Several implementation parameters are explicit. The teachers are RemoteCLIP-ViT-L/14 and DINOv3-ViT-L/16; the student is Swin Transformer-Tiny; Stage I runs for 15,000 iterations with batch size 8 and AdamW; the experiments use 8 NVIDIA A100 GPUs. The paper does not specify the learning rate, weight decay, the exact values of $\lambda_{\text{cosine}}$ and $\lambda_{\text{attn}}$, the initialization of $\mu$, or the precise Stage II loss weight, and these omissions matter for full reimplementation (Xu et al., 23 May 2026).

3. Visual prior distillation and teacher fusion

A central claim of DisDop is that RemoteCLIP and DINOv3 provide complementary visual priors. RemoteCLIP is strong in cross-modal alignment and thus useful for open-vocabulary semantics, but image-level contrastive training leaves it weaker at local feature extraction. DINOv3 is strong in local spatial representation, preserving semantic consistency across patches, but does not natively supply text alignment. DisDop fuses the two before distilling them into the student backbone (Xu et al., 23 May 2026).

For an input image, the framework extracts patch/token features

$\mathbf{F}_{\text{CLIP}}$

from RemoteCLIP and

$\mathbf{F}_{\text{DINO}}$

from DINOv3. It then computes a DINOv3 self-similarity matrix

$$\mathbf{S} = \frac{\mathbf{F}_{\text{DINO}} \cdot \mathbf{F}_{\text{DINO}}^T}{\|\mathbf{F}_{\text{DINO}}\| \cdot \|\mathbf{F}_{\text{DINO}}^T\|} \in \mathbb{R}^{N \times N},$$

where $\mathcal{L}_{\text{det}}$0 is the number of visual patch tokens. This matrix captures patchwise semantic and spatial relations and is used to structure the RemoteCLIP features (Xu et al., 23 May 2026).

A distinctive component is semantic-adaptive outlier detection (SAOD). For each patch $\mathcal{L}_{\text{det}}$1, the method defines a DINO-guided semantic neighborhood

$\mathcal{L}_{\text{det}}$2

where $\mathcal{L}_{\text{det}}$3 is a similarity threshold. Within this neighborhood, Local Outlier Factor (LOF) is applied to the RemoteCLIP features, producing a context-dependent outlier mask

$\mathcal{L}_{\text{det}}$4

with $\mathcal{L}_{\text{det}}$5 indicating that patch $\mathcal{L}_{\text{det}}$6 is an outlier relative to patch $\mathcal{L}_{\text{det}}$7’s neighborhood. The paper stresses that outlier status is relative rather than global: a patch may be anomalous for one semantic center and normal for another (Xu et al., 23 May 2026).

After filtering weak or anomalous relations, attention weights are defined by

$\mathcal{L}_{\text{det}}$8

when

$\mathcal{L}_{\text{det}}$9

and $f_c^i$0 otherwise. The fused teacher feature is then

$f_c^i$1

In effect, DINOv3 supplies the affinity structure and RemoteCLIP supplies the semantic features, yielding a teacher representation intended to reinforce intra-object consistency while preserving open-vocabulary semantics (Xu et al., 23 May 2026).

The distilled student is trained with two losses. The cosine similarity loss is

$f_c^i$2

and the attention distillation loss is

$f_c^i$3

where

$f_c^i$4

with $f_c^i$5 and $f_c^i$6. The total backbone distillation objective is

$f_c^i$7

The specific values of $f_c^i$8 and $f_c^i$9 are not reported in the provided text (Xu et al., 23 May 2026).

4. Textual priors, semantic geometry, and contextual enhancement

The paper argues that pointwise alignment between detector features and visual targets is insufficient for open-vocabulary generalization because unseen-category recognition depends on preserving semantic topology across categories. A baseline pointwise distillation objective is written as

$f_{eh}^j$0

where $f_{eh}^j$1 is the detector output category feature for ROI $f_{eh}^j$2 and $f_{eh}^j$3 is the RemoteCLIP visual feature for ROI $f_{eh}^j$4. The paper criticizes this form because it handles small objects poorly and ignores inter-category relations (Xu et al., 23 May 2026).

DisDop instead uses textual prior distillation based on pairwise relationships among RemoteCLIP text embeddings $f_{eh}^j$5. The core objective is

$f_{eh}^j$6

Here $f_{eh}^j$7 is the detector output feature for instance $f_{eh}^j$8, $f_{eh}^j$9 is an enhanced visual feature for instance $\lambda_{\text{cosine}}$0, and $\lambda_{\text{cosine}}$1 are text embeddings from RemoteCLIP’s text encoder. The semantic relation modeling is therefore implicit in pairwise cosine distances rather than in an explicit graph or prompt-ensemble mechanism. The paper gives the example that “bridge” and “ship” should be closer than “bridge” and “windmill,” illustrating how textual geometry supervises representation geometry (Xu et al., 23 May 2026).

The contextual prior is introduced because aerial images often contain tiny objects whose local ROI crops do not carry enough semantic evidence. The paper argues that global scene context can disambiguate such cases. It therefore defines an enhanced visual target

$\lambda_{\text{cosine}}$2

where $\lambda_{\text{cosine}}$3 is the local visual feature from the ROI crop, $\lambda_{\text{cosine}}$4 is a global visual feature extracted from the entire image containing the ROI, $\lambda_{\text{cosine}}$5 is a learnable balancing parameter, and $\lambda_{\text{cosine}}$6 is layer normalization. In the framework description, this contextual prior modifies the training target used in relational distillation rather than introducing a separate inference-time context head (Xu et al., 23 May 2026).

This design indicates that textual and contextual priors are coupled in Stage II: the model is not only encouraged to match local visual targets but to organize its category features according to a language-derived geometry, while that geometry is anchored to supervision targets enhanced with global scene context. A plausible implication is that DisDop is particularly tuned to the two central aerial-detection failure modes named in the paper: unseen-category semantics and small-object ambiguity (Xu et al., 23 May 2026).

5. Training data, benchmarks, and evaluation protocol

DisDop uses LAE-1M for training in both stages: unlabeled LAE-1M in Stage I and labeled LAE-1M in Stage II. Evaluation is conducted on three aerial benchmarks: DIOR, DOTAv2.0, and LAE-80C. The paper reports open-set detection results in Table 1 and closed-set fine-tuning results in Table 2. For DOTAv2.0, only horizontal bounding box detection is evaluated (Xu et al., 23 May 2026).

The reported metrics are $\lambda_{\text{cosine}}$7 and $\lambda_{\text{cosine}}$8. The supplied text does not specify seen/unseen class splits, whether the setting is generalized zero-shot versus strict unseen-only, or the exact prompt construction used for category texts. The paper excerpt also does not report per-seen and per-unseen AP values. These omissions are material because they limit full protocol reconstruction and make the benchmark definition less transparent than the model formulation itself (Xu et al., 23 May 2026).

The hyperparameters explicitly reported include the Stage I teacher-fusion thresholds

$\lambda_{\text{cosine}}$9

and the attention-distillation temperatures

$\lambda_{\text{attn}}$0

Best-performing configuration choices vary by setting, but the paper excerpt states that the detector is built on LAE, the student backbone is Swin Transformer-Tiny, and teacher models are RemoteCLIP-ViT-L/14 and DINOv3-ViT-L/16 (Xu et al., 23 May 2026).

6. Empirical performance, ablations, and limitations

The main open-set detection results show consistent gains over prior baselines. On DIOR, the paper reports: GLIP + DVC: 82.8 $\lambda_{\text{attn}}$1, GroundingDINO + DVC: 83.6, LAE-DINO: 85.5, and DisDop: 87.2. On DOTAv2.0, the reported numbers are 43.0, 46.0, 46.8, and 47.5 $\lambda_{\text{attn}}$2, respectively. On LAE-80C, they are 16.5, 17.7, 20.2, and 22.5 $\lambda_{\text{attn}}$3, with the largest relative gain occurring on LAE-80C (Xu et al., 23 May 2026).

In closed-set fine-tuning, DisDop also improves over the strongest listed baseline. On DIOR, the sequence is GLIP-FT: 88.9, GroundingDINO-FT: 91.1, LAE-DINO-FT: 92.2, and DisDop-FT: 92.6. On DOTAv2.0, it is 51.5, 55.1, 57.9, and 58.6, respectively. These numbers indicate that the domain-prior distillation strategy is not confined to open-set transfer, but also benefits conventional fine-tuning (Xu et al., 23 May 2026).

Benchmark	Baseline	DisDop
DIOR open-set $\lambda_{\text{attn}}$4	LAE-DINO 85.5	87.2
DOTAv2.0 open-set $\lambda_{\text{attn}}$5	LAE-DINO 46.8	47.5
LAE-80C open-set $\lambda_{\text{attn}}$6	LAE-DINO 20.2	22.5
DIOR closed-set $\lambda_{\text{attn}}$7	LAE-DINO-FT 92.2	92.6
DOTAv2.0 closed-set $\lambda_{\text{attn}}$8	LAE-DINO-FT 57.9	58.6

The ablation studies are designed to isolate the effect of the framework’s major components. With LAE-DINO as baseline, adding only VPD improves DIOR from 85.5 to 86.6, DOTAv2.0 from 46.8 to 47.2, and LAE-80C from 20.2 to 21.5. Adding only TCPD yields 86.2, 47.1, and 20.9. Combining both gives the full 87.2, 47.5, and 22.5. This shows that both visual and textual/contextual distillation contribute, while the visual branch provides the larger single gain in the reported table (Xu et al., 23 May 2026).

Teacher-fusion ablations further support the role of SAOD. The paper reports, for the visual-distillation stage, w/o Outlier Filtering: 86.4 / 47.1 / 21.3, Global LOF: 86.3 / 47.0 / 21.2, and SAOD (Ours): 86.6 / 47.2 / 21.5 on DIOR, DOTAv2.0, and LAE-80C, respectively. A separate ablation on relational-distillation scope gives Point-Wise: 85.8 / 46.8 / 20.4, Within-Image: 86.1 / 46.9 / 20.7, and Within-Batch (Ours): 86.2 / 47.1 / 20.9, supporting the paper’s claim that batch-level semantic relations are denser and more useful than within-image ones in aerial scenes (Xu et al., 23 May 2026).

Several limitations remain explicit. The supplied text states that the paper does not provide exact layer mappings between teacher and student, exact projection/alignment head architectures, the full weighted Stage II objective, prompt templates for text embeddings, the exact inference-time classification formulation, or seen/unseen split definitions for the open-vocabulary benchmarks. These omissions do not alter the conceptual identity of DisDop, but they do constrain reproducibility from the paper text alone. A plausible implication is that code consultation would be necessary for faithful reimplementation, particularly in Stage II where the textual and contextual priors interface with LAE’s detector-specific components (Xu et al., 23 May 2026).