Bar-JEPA: Extracting Values from Bar Chart with Joint-Embedding Predictive Architecture

Extracting raw data from bar charts is essential for accessibility and automated analysis, yet it remains a difficult task due to the scarcity of high-quality, labeled real-world data. The authors investigate whether a self-supervised approach, specifically the Joint-Embedding Predictive Architecture (JEPA), can learn semantically rich features from unlabeled charts to improve the efficiency and accuracy of numerical value recovery compared to traditional supervised methods.

The researchers propose a pipeline that leverages I-JEPA, a vision-based JEPA model, as a feature extractor. They modify the architecture to handle variable-resolution inputs, allowing the model to process charts of varying aspect ratios without losing spatial information. The pipeline consists of two main stages: self-supervised pretraining on a large synthetic dataset of 100k bar charts, followed by training a lightweight decoder to regress heatmaps for bar and tick locations. The decoder is designed to be simple and computationally efficient, focusing on the quality of the latent features provided by the JEPA encoder.

Bar-JEPA demonstrates that self-supervised pretraining on synthetic data effectively captures the structural and semantic nuances of bar charts. By utilizing latent features rather than pixel-level reconstruction, the model achieves robust performance in de-rendering charts. The authors show that this approach is more data-efficient than end-to-end supervised baselines, which often struggle with the diversity and limited availability of real-world annotated charts. The inclusion of variable-resolution support further enhances the model's ability to generalize across different chart layouts.

This work addresses a significant bottleneck in document understanding: the reliance on massive, manually labeled datasets for chart analysis. By demonstrating that JEPA can learn powerful representations from synthetic data, the authors provide a scalable path for building more accessible and data-efficient visualization analysis tools. This approach is particularly relevant for researchers working in domains where chart data is abundant but annotated ground truth is rare.