Search papers, labs, and topics across Lattice.
This paper introduces Geometry-Aware Pointer (GAP) Loss, a novel modification to the loss function used in Table Structure Recognition (TSR) that emphasizes spatial locality by reweighting negative candidates based on their proximity to ground truth. The authors identify that a significant portion of errors in pointer networks occurs between adjacent cells, and GAP Loss effectively addresses this by applying inverse distance weighting to enhance gradient flow for nearby cells. Extensive evaluations on datasets like PubTabNet and SynthTabNet reveal that GAP Loss significantly reduces adjacent-cell errors, achieving state-of-the-art performance in TSR tasks.
Reweighting loss based on spatial proximity can dramatically reduce adjacent-cell errors in table structure recognition, leading to breakthrough performance improvements.
Table Structure Recognition (TSR) using a pointer network achieves impressive results by predicting HTML sequences while aligning tags to detected text (or cell) regions. However, our analysis reveals that when pointer networks fail, 79.6% of errors occur between spatially adjacent cells (Manhattan distance <= 2). Despite this, standard cross-entropy loss weights all negative candidates equally. In this work, we propose Geometry-Aware Pointer (GAP) Loss, which reweights the cross-entropy objective based on spatial proximity to ground truth. By applying inverse distance weighting, GAP focuses gradient flow where the model struggles most: immediate neighbors receive stronger gradients than distant cells. Our approach requires only a straightforward modification to the loss computation, maintaining the same model architecture with zero additional inference cost. Extensive experiments on PubTabNet and SynthTabNet demonstrate that GAP consistently reduces adjacent-cell errors, achieving new state-of-the-art performance. Our findings suggest that incorporating geometric inductive biases at the loss level provides a simple yet effective approach to robust TSR. Our code is available at https://github.com/teamreboott/GAP
