Learning Human Perspective in Line Drawings from Single Sketches

2025-04-04

AI Summary

This work addresses the challenge of modeling implicit human-centric perspective biases in single freehand sketches. We propose the first few-shot learning framework capable of automatically inferring and generalizing non-Euclidean, artist-specific perspective rules from a single line drawing. Methodologically, we align strokes with analytic camera projections via geometric contour matching, enabling a differentiable, spatially continuous local perspective correction function. To ensure cross-view consistency, we introduce viewpoint-neighborhood data augmentation and a single-sample meta-learning mechanism. Compared to conventional rigid perspective models, our method produces rendered sketches that better preserve artistic style and significantly improve visual plausibility and interactive naturalness in sketch-based modeling and non-photorealistic rendering tasks. Experimental results validate both the effectiveness and practical utility of human-centered perspective modeling.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Learning with Manifolds
• Cognitive Modeling & Cognitive Systems: Computational Creativity

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Artist-drawn sketches only loosely conform to analytical models of perspective projection. This deviation of human-drawn perspective from analytical perspective models is persistent and well known, but has yet to be algorithmically replicated or even well understood. Capturing human perspective can benefit many computer graphics applications, including sketch-based modeling and non-photorealistic rendering. We propose the first dedicated method for learning and replicating human perspective. A core challenge in learning this perspective is the lack of suitable large-scale data, as well as the heterogeneity of human drawing choices. We overcome the data paucity by learning, in a one-shot setup, from a single artist sketch of a given 3D shape and a best matching analytical camera view of the same shape. We match the contours of the depicted shape in this view to corresponding artist strokes. We then learn a spatially continuous local perspective deviation function that modifies the camera perspective projecting the contours to their corresponding strokes while retaining key geometric properties that artists strive to preserve when depicting 3D content. We leverage the observation that artists employ similar perspectives when depicting shapes from slightly different view angles to algorithmically augment our training data. First, we use the perspective function learned from the single example to generate more human-like contour renders from nearby views; then, we pair these renders with the analytical camera contours from these views and use these pairs as additional training data. The resulting learned perspective functions are well aligned with the training sketch perspectives and are consistent across views. We compare our results to potential alternatives, demonstrating the superiority of the proposed approach, and showcasing applications that benefit from learned human perspective.

Problem

• Learning human perspective deviations from analytical models in sketches
• Overcoming data scarcity by one-shot learning from single artist sketches
• Generating human-like contour renders for sketch-based applications

Innovation

• Learning human perspective from single sketches
• One-shot setup with analytical camera matching
• Spatially continuous local perspective deviation function