RegionCL: Region Swapping for Contrastive Learning

Updated 7 June 2026

RegionCL is a self-supervised learning framework that extends contrastive learning by swapping image regions to create paste and canvas views for spatially-aware feature learning.
It employs a batch-wise region swapping operation that exchanges image patches to generate structured positive and negative pairs, boosting performance across various tasks.
Experimental results show that RegionCL achieves notable gains over conventional SSL methods, improving ImageNet linear probing, COCO detection, and Cityscapes segmentation.

RegionCL is a self-supervised learning (SSL) framework that augments contrastive learning with region swapping as a core pretext task. While conventional SSL approaches typically leverage random crops to generate positive pairs for maximizing mutual information between two augmented image views, RegionCL systematically utilizes both the cropped (paste) and the complementary (canvas) regions of source images. Central to RegionCL is a batch-wise region swapping operation: it forms composite images by exchanging random patches between images and treats the resulting subregions as additional, structure-aware contrastive views. This enables the model to learn discriminative representations that are sensitive to the internal composition of visual instances, thereby improving transferability to classification, detection, and segmentation tasks (Xu et al., 2021).

1. Methodological Foundations and Region Swapping

RegionCL expands the scope of contrastive data augmentation beyond standard random cropping by introducing explicit region swaps. For each image $x_i$ in a batch, two global views $x^q_i$ and $x^k_i$ are generated using standard augmentations (random crop, color jitter, flip, blur). From $x^q_i$ , a rectangular patch $x^p$ (the "paste view") is randomly cropped, with size $(r_w, r_h) = (c_w \cdot R, c_h \cdot R)$ where $(c_w, c_h)$ are uniformly sampled integers in $[\mathcal C_L, \mathcal C_U]$ and $R$ is the encoder's downsampling rate (e.g., 32 for ResNet-50).

This patch is swapped with a patch from another randomly selected image in the batch. The composite images, $x_A' = \text{paste}(p_B, c_A)$ and $x^q_i$ 0, each contain a paste view (from another image) and a canvas view (remaining content from the original image). This augmentation strategy generates abundant, structured pairs: paste/canvas pairs from the same composite (negative), as well as canvas/original and paste/original (positive) pairs, and hard negatives across different images.

2. Loss Design and Mathematical Formulation

RegionCL builds on classic contrastive objectives and extends them with region-level loss terms. The total loss combines:

Instance-level InfoNCE:

$x^q_i$ 1

where $x^q_i$ 2 is the projection of the query (from online net), $x^q_i$ 3 the positive key (from momentum net), and $x^q_i$ 4 is the queue size.

Region-level RegionCL Loss:

$x^q_i$ 5

where $x^q_i$ 6 are the masked-pooled features for paste and canvas, and $x^q_i$ 7 denotes stop-gradient to enforce hard negatives.

The full loss is:

$x^q_i$ 8

This framework can be integrated into MoCo v2, DenseCL, and SimSiam with minimal architectural changes by adding branch heads for the new composite views and corresponding mask pooling to extract region-wise features.

3. Experimental Evaluation and Empirical Results

RegionCL achieves consistent improvements across benchmarks and SSL baselines. Representative results include:

ImageNet Linear Probe (Top-1/Top-5, Epochs):
- MoCo v2: 67.5 / 88.2 (200)
- RegionCL-M: 69.4 / 89.6 (200)
- DenseCL: 63.6 / 85.5 (200)
- RegionCL-D: 68.5 / 89.0 (200)
- SimSiam: 68.1 / 88.2 (100)
- RegionCL-S: 71.3 / 90.4 (100)
COCO Detection (AP/AP50/AP75):
- MoCo v2: 38.9/59.4/42.4 (box), 35.5/56.5/38.1 (mask)
- RegionCL-M: 40.1/60.7/43.9, 36.3/57.7/39.0
- DenseCL: 39.1/59.4/42.5, 35.5/56.4/38.0
- RegionCL-D: 40.4/61.3/44.2, 36.7/58.2/39.4
Cityscapes Segmentation (AP/75, mIoU 40K/80K):
- MoCo v2: 33.9/60.8, 77.8/78.6
- RegionCL-M: 34.9/62.5, 78.1/79.0

Across all tasks, introducing region-level contrastive augmentation outperforms the corresponding vanilla SSL method, with the most pronounced gains on object detection and segmentation, where structured, spatially-aware representations are particularly beneficial (Xu et al., 2021).

4. Ablation Studies and Hyperparameter Effects

RegionCL includes thorough ablations:

Patch size bounds ( $x^q_i$ 9, $x^k_i$ 0): optimal results are observed with $x^k_i$ 1, $x^k_i$ 2, corresponding to patch sizes in $x^k_i$ 3 pixels for a typical $x^k_i$ 4.
View composition: Using both paste and canvas regions in the loss (plus their negatives) yields the best performance; using only one region provides notable but smaller gains.
Swapping vs. Zero-Fill: Region swapping (as opposed to erasing or zero-filling the patch) further improves accuracy by creating harder negatives.

The implementation is efficient: coordinates are aligned to the feature map grid, and batch-wise swapping is performed for computational efficiency. All region and instance features share the same queue in MoCo-style approaches.

RegionCL is architecturally agnostic and can be integrated into varied contrastive and Siamese SSL paradigms with minor modifications. Unlike pixel-level or image-level contrastive frameworks, RegionCL is explicitly structured to leverage both the foreground ("paste") and complementary background ("canvas") regions, providing a richer set of contrastive pairs. This approach is distinct from prior works that employ region masking, erasure, or CutMix-style augmentation; RegionCL establishes structured intra-image negatives by explicit swapping (Xu et al., 2021).

Other contemporary region-level SSL and representation learning efforts—such as RegionCLIP (Zhong et al., 2021), RC²L (Zhang et al., 2022), and RICE (Xie et al., 26 Jul 2025)—expand region-level supervision or contrastive learning to language-image, semantic segmentation, or clustering tasks, respectively. RegionCL's contribution is specifically the design of the region swapping pretext, its contrastive pairing logic, and demonstration of its broad effectiveness in image recognition pipelines.

6. Implementation and Practical Guidance

RegionCL is published with detailed code and default configurations (https://github.com/Annbless/RegionCL). Key practical recommendations include:

Patch coordinates should align to the downsampling grid for precise mask-pooling.
Use a single dictionary queue for all features in MoCo-style methods.
Perform region swap batch-wise for all images in a mini-batch to maximize computational efficiency.
Recommended hyperparameters: $x^k_i$ 5, $x^k_i$ 6; temperature $x^k_i$ 7; queue size $x^k_i$ 8; batch size 1024 (MoCo/DenseCL), 512 (SimSiam).
Adopt standard augmentations as in MoCo v2 for input diversity.

With these design and engineering choices, RegionCL consistently gives state-of-the-art or competitive results in self-supervised visual representation learning frameworks (Xu et al., 2021).

References

RegionCL: "RegionCL: Can Simple Region Swapping Contribute to Contrastive Learning?" (Xu et al., 2021)
RegionCLIP: "RegionCLIP: Region-based Language-Image Pretraining" (Zhong et al., 2021)
RC²L: "Region-level Contrastive and Consistency Learning for Semi-Supervised Semantic Segmentation" (Zhang et al., 2022)
RICE: "Region-based Cluster Discrimination for Visual Representation Learning" (Xie et al., 26 Jul 2025)