Control Copy-Paste for Remote Sensing FSOD

• Control Copy-Paste is a controllable diffusion-based data augmentation method designed for few-shot object detection in optical remote sensing, addressing overfitting issues.
• It employs a conditional diffusion model that harmoniously integrates target objects with varied contextual backgrounds while handling extreme aspect ratios with an orientation alignment strategy.
• Evaluations on DIOR demonstrate that the method improves FSOD performance by an average of 10.76%, highlighting its practical significance in remote sensing applications.

Control Copy-Paste is a controllable diffusion-based data augmentation method for optical remote sensing few-shot object detection (FSOD). In this setting, the objective is to detect rare objects from only a few annotated bounding boxes, and the method is motivated by the claim that overfitting arises not only from limited novel-instance diversity but also from monotonous contextual backgrounds. The approach uses a conditional diffusion model to seamlessly inject few-shot novel objects into images with diverse contexts, introduces an orientation alignment strategy for varying aspect ratios, and reports an average improvement of 10.76% on the DIOR benchmark (Liu et al., 29 Jul 2025).

1. Problem setting and conceptual motivation

Control Copy-Paste is formulated within the standard two-stage FSOD protocol. The category set is partitioned into base classes $C_b$ and novel classes $C_n$, with

$C_b \cap C_n = \emptyset .$

The detector is first trained on base classes with sufficient labeled data and then fine-tuned on novel classes using only $K$-shot annotated examples. This protocol is conventional in FSOD, but the paper argues that remote sensing introduces a distinctive failure mode: detectors can overfit not only to the small number of novel objects, but also to the small number of scene contexts in which those objects appear (Liu et al., 29 Jul 2025).

A common simplification is to treat few-shot performance as a problem of object scarcity alone. Control Copy-Paste rejects that simplification explicitly. The paper states that naively increasing the diversity of objects is insufficient, because surrounding contexts also play a crucial role in object detection, and in cases where object diversity is sufficient, the detector tends to overfit to monotonous contexts. In remote sensing, this claim is especially consequential because the images are large, the backgrounds are information-rich, and the background can dominate the learned representation. The proposed augmentation strategy therefore targets contextual diversity as a first-class variable rather than as an incidental by-product of pasting (Liu et al., 29 Jul 2025).

2. Context diversification as the central control mechanism

The method is described as a pipeline that takes a few-shot reference image containing a target object, a context image, and the location or mask where the object should be integrated, and then outputs a synthetic training image. Its purpose is not merely to place more instances into the dataset, but to generate many versions of the same few-shot novel object in different environments. The intended effect is that the detector sees the same object under more varied contexts, becomes less sensitive to overfitted background patterns, and learns more robust object-context separation (Liu et al., 29 Jul 2025).

This emphasis distinguishes the method from copy-paste schemes that are primarily instance-centric. The paper states that standard copy-paste improves context diversity, but because it relies on box-level annotation, it cannot fully decouple object and background. Control Copy-Paste is presented as a response to that limitation: instead of crude insertion, it uses diffusion-based controlled integration so that the inserted object is harmoniously integrated with the destination scene. The paper further claims that, relative to standard copy-paste, the proposed method can achieve up to 3.81% additional improvement because it decouples context and instance more effectively (Liu et al., 29 Jul 2025).

3. Conditional diffusion pipeline

The generative core is a class-agnostic conditional diffusion model inspired by AnyDoor and Stable Diffusion. The reverse process is written as

$$p(z_{t-1} \mid z_t, x_c, m_c, x_r) = UNet(z_t, x_c \odot m_c, x_r),$$

where $z_t$ is the latent at diffusion step $t$, $x_c$ denotes context-image features, $m_c$ is the context mask, and $x_r$ is the reference-image representation. In plain terms, the U-Net predicts the denoised latent conditioned jointly on the masked context and on the reference object, so the model learns how to blend the target object into the background realistically (Liu et al., 29 Jul 2025).

The reference pathway has two conditioning streams. First, coarse-grained features are obtained from DINOv2. The reference image is resized to

$C_n$0

with the longer side resized to 224 and the shorter side padded before resizing in order to preserve aspect ratio. The DINOv2 representation is encoded into

$C_n$1

and then mapped to the U-Net embedding space through a linear layer. Second, fine-grained guidance is obtained by high-pass filtering, which captures edges and detailed shape cues. The paper’s synthesis logic is therefore explicitly dual-scale: DINOv2 supplies robust semantic structure, while high-pass filtering supplies local geometric detail (Liu et al., 29 Jul 2025).

The method is also described as approach-agnostic. This means that the augmentation module is not tied to a single FSOD detector; instead, synthetic images generated by the diffusion model are added during novel-class fine-tuning. The paper also notes the use of Gradual RPN to improve localization accuracy in the training pipeline (Liu et al., 29 Jul 2025).

4. Orientation alignment and reference transformation

A specific technical problem in remote sensing is extreme aspect ratio variation. The paper highlights that many remote sensing objects are elongated, and that mismatch between the pasted object and the target slot can cause visible integration distortion. Control Copy-Paste addresses this with an orientation alignment strategy that aligns the long edge of the reference instance with the long edge of the target region (Liu et al., 29 Jul 2025).

The alignment rule is defined as

$C_n$2

where $C_n$3 and $C_n$4 are the aspect ratios of the reference instance and the target area. If the reference and target have opposite orientation tendencies, the reference is rotated by $C_n$5; otherwise it is left unchanged. The paper identifies this as especially important for highly elongated objects such as train stations, for which aspect-ratio mismatch can otherwise degrade performance (Liu et al., 29 Jul 2025).

The full reference transformation is written as

$C_n$6

where $C_n$7 is orientation alignment, $C_n$8 is high-pass filtering, and $C_n$9 denotes affine transformations such as rotation, mirroring, and flipping. In operational terms, the reference object is first aligned, then filtered to emphasize fine-grained detail, and then augmented to increase viewpoint diversity. This sequence clarifies that control in Control Copy-Paste is not limited to semantic conditioning; it also includes explicit geometric normalization and transformation of the pasted instance (Liu et al., 29 Jul 2025).

5. Experimental protocol and reported performance

The paper evaluates the method on DIOR, a remote sensing object detection dataset with 20 total classes, divided into 15 base classes and 5 novel classes. The novel classes are airplane, baseball field, tennis court, train station, and windmill. Few-shot evaluation is conducted under 3-shot, 5-shot, 10-shot, and 20-shot settings, using $C_b \cap C_n = \emptyset .$0 as the metric. Training is performed on DIOR trainval, evaluation on DIOR test, and context images are sampled from DIOR scenes that do not contain novel instances; synthetic data is then combined with the few-shot training set to avoid hallucination (Liu et al., 29 Jul 2025).

Two FSOD baselines are reported: FSCE and MSOCL. For MSOCL + Ours, the paper reports 38.79 mAP at 3-shot, 39.99 at 5-shot, 41.70 at 10-shot, and 44.46 at 20-shot. For FSCE + Ours, it reports 32.18, 33.21, 34.52, and 35.78 mAP, respectively. These numbers are said to consistently outperform both the original baselines and baseline-plus-standard-copy-paste variants. The headline claim is that Control Copy-Paste improves detection performance by an average of 10.76% on the novel classes (Liu et al., 29 Jul 2025).

The reported qualitative interpretation is equally important. The paper states that increasing context diversity can substantially improve FSOD performance, sometimes making a low-shot setup competitive with a higher-shot one. It also states that once context diversity becomes sufficient, object diversity may again become the limiting factor. This makes the empirical argument explicitly bidirectional: object diversity and context diversity are complementary, and neither alone is represented as a complete solution (Liu et al., 29 Jul 2025).

6. Relation to broader copy-paste research and stated limitations

Within the broader copy-paste literature, Control Copy-Paste occupies a specific position. X-Paste reframed copy-paste as a scalable data acquisition problem for instance segmentation by using CLIP and StableDiffusion to acquire large pools of pasteable instances, reporting +2.6 box AP and +2.1 mask AP overall on LVIS, and +6.8 box AP and +6.5 mask AP on rare classes (Zhao et al., 2022). Context-Aware Copy-Paste argued that standard augmentation often neglects semantic coherence and used BLIP, YOLO, and SAM to match pasted content to scene context, reporting improvements over random copy-paste across classification, segmentation, and detection tasks (Guo, 2024). In crowded-scene learning, Occlusion Copy & Paste and crowdedness-oriented copy-paste used controlled overlap to synthesize same-class occlusions and crowded groups for human and pedestrian detection (Ling et al., 2022, Deng et al., 2022). This suggests a broader shift in which copy-paste is increasingly treated not as literal cut-and-paste, but as a controlled data-generation mechanism tailored to the failure mode of a target task.

Control Copy-Paste extends that trajectory into remote sensing FSOD by making contextual diversity the main augmentation target and by using conditional diffusion rather than direct compositing. A plausible implication is that its central contribution is less the existence of synthetic objects than the ability to harmonize few-shot objects with diverse backgrounds while handling extreme aspect ratios. The paper nonetheless states clear constraints. Extreme aspect-ratio objects remain challenging, with train station noted as a category where performance still drops somewhat. The method assumes the availability of a reference object, a suitable context image, and a target integration region. It also relies on context images selected so as not to contain novel instances, which the paper presents as a controlled experimental setting that helps prevent hallucination (Liu et al., 29 Jul 2025).