How Context-Contrast Features and Gated Multi‑Scale Fusion Boost Scene Segmentation

This paper, selected among 18 Alibaba papers at IEEE CVPR‑2018, addresses the scene segmentation problem, which requires pixel‑level classification and benefits from contextual information and multi‑scale feature fusion.

Authors: Henghui Ding, Xudong Jiang, Bing Shuai, Ai Qun Liu (School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore) and Gang Wang (Alibaba AI Labs, Hangzhou, China).

The authors first propose a context‑contrast local feature that leverages rich contextual information while contrasting it with local details, thereby highlighting discriminative local cues and improving segmentation of both salient and non‑salient objects as well as background regions.

To further enhance performance, the paper introduces a gated multi‑scale fusion mechanism. Unlike traditional skip‑connection fusion that simply adds features, the gated fusion dynamically weights multi‑scale features based on the input image’s characteristics, allowing the network to adaptively select the most appropriate receptive field for each pixel.

The gating values are generated by the proposed network and vary with each image, controlling the flow of information across scales and strengthening the model’s ability to handle objects of diverse sizes.

Experiments on three benchmark datasets—Pascal Context, SUN‑RGBD, and COCO Stuff—show that the proposed method achieves the highest reported scene‑segmentation performance, particularly improving the segmentation of inconspicuous objects and background regions while maintaining strong adaptability to multi‑scale objects.

Qualitative results (Figure 4) illustrate the clear improvement in segmenting both non‑salient objects and background compared with previous approaches.