Wide2Long: Learning Lens Compression and Perspective Adjustment for Wide-Angle to Telephoto Translation

Soumyadipta Banerjee, Jiaul H. Paik, Debashis SenICCV 2025

TL;DR

This paper introduces "Wide2Long" (W2L), a novel image-to-image translation pipeline that simulates the effect of a telephoto lens from a wide-angle lens image. This allows for achieving lens compression and perspective adjustment without needing specialized hardware, making professional-looking shots more accessible.

What Problem Does This Paper Address?

Professional photographers understand that different lenses capture subjects in distinct ways, primarily due to their focal lengths. Wide-angle lenses (short focal lengths) capture broad scenes, ideal for landscapes, while telephoto lenses (long focal lengths) offer a narrower field of view, better suited for portraits and distant subjects. Smartphones, due to size and cost constraints, typically only feature wide-angle lenses. While some high-end phones offer multiple lenses, switching between them can be cumbersome. This paper addresses the challenge of digitally transforming images captured with a wide-angle lens to appear as if they were taken with a telephoto lens, specifically by simulating "lens compression" and perspective adjustments. This is distinct from simple digital zoom, which only magnifies pixels without altering perspective or depth perception.

Methods

The Wide2Long (W2L) pipeline is a multi-stage approach designed to achieve this translation:

  • Depth-based Layering (DBL): The pipeline first estimates the depth map of the input wide-angle image using a pre-trained MiDaS model. This depth information is then used to segment the image into multiple "depth layers."
  • Layer-wise Inpainting (LI): Since the depth layering can create occluded regions (areas hidden behind foreground objects), an in-painting technique is employed. The authors leverage a fast Fourier convolution (FFC) based in-painting model to fill these gaps, considering global image context. To improve in-painting quality, they introduce a "composite context layer" and a dilated mask to incorporate information from surrounding layers.
  • Inter-layer Redundancy Reduction (ILRR): An issue that can arise is redundant visual information appearing in adjacent depth layers after in-painting. This module uses a Structural Similarity Index Measure (SSIM) to identify and reduce this redundancy, preventing distortions in the final output.
  • Layer Scaling and Flattening (LSF): This final stage scales each processed depth layer to achieve the desired lens compression effect. The core idea is that objects at a specific distance from the camera should maintain their size in both the original wide-angle and the translated telephoto image. A learnable parameter ($\gamma$) controls the magnification/demagnification of layers based on their depth relative to a user-specified "subject" pixel. The scaled layers are then fused to produce the final telephoto-like image.

Crucially, the parameters for these modules (specifically $\lambda$, $\gamma$, $T$, and $\alpha$) are learned by a convolutional neural network called "Image2Param" (I2P), trained on image-ground truth pairs.

Datasets

  • New Dataset: The authors created a new dataset consisting of image pairs captured with both wide-angle and telephoto lenses at different distances to maintain subject size. This dataset includes 500 pairs for training and 100 for testing, with each image accompanied by coordinates to identify a subject for scaling.
  • Caltech Multi-Distance Portraits (CMDP) dataset: Used for evaluating the pipeline's performance on portrait perspective distortion correction.

Key Insights and Ideas

  • Depth as a Key to Translation: The paper highlights that understanding and segmenting an image by depth is fundamental to simulating lens compression. Different depth layers require different processing.
  • In-painting is Crucial, but Needs Context: While in-painting is essential for filling occluded regions, simply in-painting each layer independently leads to artifacts. The proposed "composite context layer" and dilation strategy significantly improve results by leveraging surrounding information.
  • Redundancy Reduction Prevents Distortions: The ILRR module is vital for ensuring a smooth and artifact-free transition between layers, preventing visual inconsistencies that can arise from overlapping in-painted content.
  • Learned Scaling for Realistic Compression: The use of a learned scaling factor ($\gamma$) based on a user-defined subject is key to achieving realistic lens compression, where the subject's size remains consistent while the background and foreground are appropriately scaled.
  • Versatility Beyond Compression: The pipeline demonstrates effectiveness not only for general lens compression but also for a specific application: correcting portrait perspective distortion, which is a form of lens compression applied to faces.

Key Results

  • Quantitative Performance: The W2L pipeline achieved state-of-the-art or comparable results on their custom dataset and the CMDP dataset, outperforming naive U-Net baselines and showing competitive results against other related methods like Shih et al. [15] for Dolly-Zoom. For instance, on their dataset, W2L achieved an SSIM of 0.768, PSNR of 19.873, and LPIPS of 0.139.
  • Qualitative Performance: The generated images visually demonstrate successful lens compression and perspective adjustment, producing outputs that convincingly resemble telephoto shots. The paper also shows qualitative results for higher orders of lens compression (4x and 8x).
  • Ablation Studies: Ablation studies confirmed the importance of each component. The in-painting stage was found to have the most significant impact on performance. Removing redundancy reduction and dilation parameters also negatively affected results, though to a lesser extent than in-painting.

Limitations and Future Work

  • Handling New Objects: The authors acknowledge that the pipeline may struggle when the camera movement required for telephoto simulation would expose new objects or parts of objects that were not present in the original wide-angle view. The W2L estimate might not accurately represent these newly exposed areas if they are not inferable from the input.
  • Future Work: The paper suggests that while W2L performs well, further research could explore more robust handling of scenarios where new visual information is introduced during the simulated camera movement.

Summary updated

BibTeX
@inproceedings{banerjee2025wide,
  author = {Soumyadipta Banerjee and Jiaul H. Paik and Debashis Sen},
  title = {{Wide2Long: Learning Lens Compression and Perspective Adjustment for Wide-Angle to Telephoto Translation}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2025},
}
Figure