PPMN: Pixel-Phrase Matching Network for One-Stage Panoptic Narrative Grounding

TL;DR

This paper introduces PPMN, a one-stage end-to-end network for Panoptic Narrative Grounding that directly matches phrases to pixels, improving accuracy over previous two-stage methods.

Contribution

The paper proposes a novel one-stage Pixel-Phrase Matching Network with a Language-Compatible Pixel Aggregation module for better phrase-to-pixel matching in PNG tasks.

Findings

Achieves state-of-the-art performance on PNG benchmark

Outperforms two-stage methods by 4.0 absolute Average Recall

Demonstrates effective pixel-phrase correspondence modeling

Abstract

Panoptic Narrative Grounding (PNG) is an emerging task whose goal is to segment visual objects of things and stuff categories described by dense narrative captions of a still image. The previous two-stage approach first extracts segmentation region proposals by an off-the-shelf panoptic segmentation model, then conducts coarse region-phrase matching to ground the candidate regions for each noun phrase. However, the two-stage pipeline usually suffers from the performance limitation of low-quality proposals in the first stage and the loss of spatial details caused by region feature pooling, as well as complicated strategies designed for things and stuff categories separately. To alleviate these drawbacks, we propose a one-stage end-to-end Pixel-Phrase Matching Network (PPMN), which directly matches each phrase to its corresponding pixels instead of region proposals and outputs panoptic segmentation by simple combination. Thus, our model can exploit sufficient and finer cross-modal semantic correspondence from the supervision of densely annotated pixel-phrase pairs rather than sparse region-phrase pairs. In addition, we also propose a Language-Compatible Pixel Aggregation (LCPA) module to further enhance the discriminative ability of phrase features through multi-round refinement, which selects the most compatible pixels for each phrase to adaptively aggregate the corresponding visual context. Extensive experiments show that our method achieves new state-of-the-art performance on the PNG benchmark with 4.0 absolute Average Recall gains.