Corner2Net: Detecting Objects as Cascade Corners

TL;DR

The cascade corner pipeline is devised which progressively predicts the associated corner pair in two steps instead of synchronously searching two independent corners via parallel heads, and which surpasses all existing corner-based detectors by a large margin in accuracy and speed.

Abstract

The corner-based detection paradigm enjoys the potential to produce high-quality boxes. But the development is constrained by three factors: 1) Hard to match corners. Heuristic corner matching algorithms can lead to incorrect boxes, especially when similar-looking objects co-occur. 2) Poor instance context. Two separate corners preserve few instance semantics, so it is difficult to guarantee getting both two class-specific corners on the same heatmap channel. 3) Unfriendly backbone. The training cost of the hourglass network is high. Accordingly, we build a novel corner-based framework, named Corner2Net. To achieve the corner-matching-free manner, we devise the cascade corner pipeline which progressively predicts the associated corner pair in two steps instead of synchronously searching two independent corners via parallel heads. Corner2Net decouples corner localization and object classification. Both two corners are class-agnostic and the instance-specific bottom-right corner further simplifies its search space. Meanwhile, RoI features with rich semantics are extracted for classification. Popular backbones (e.g., ResNeXt) can be easily connected to Corner2Net. Experimental results on COCO show Corner2Net surpasses all existing corner-based detectors by a large margin in accuracy and speed.

Figures (4)

• Figure 1: Comparision of parallel corner detection and the proposed cascade corner detection. All existing corner-based methods fall into the parallel detection pipeline, which predicts two separate class-specific corners and relies on the corner-matching algorithm when decoding boxes. Corner2Net adopts the proposed cascade corner detection pipeline that decouples the corner localization and object classification. Corner2Net mines all objects via the class-agnostic associated corner pair which is more robust, and the instance-specific bottom-right corner further simplifies its search space. The object category is predicted using RoI features with rich instance semantics. Our Corner2Net runs in a corner-matching-free manner, and it can achieve higher accuracy and efficiency under various backbones and is not limited to the Hourglass network.
• Figure 2: The framework of the proposed Corner2Net. Corner2Net has two cascade stages to conquer the corner-based detection task. In the first stage, all class-agnostic top-left corners are located on the image-level heatmap and adjusted by offsets, and each RoI space is determined to establish an association between two cascade corners. Next, the RoI features are fed into the second stage, where the precise bottom-right corner of each specific instance is obtained by its instance-level heatmap and offset. Meanwhile, the category is predicted by a lightweight head with rich instance semantics.
• Figure 3: Qualitative detection results of CornerNet baseline and the proposed Corner2Net on MS-COCO val2017 set.
• Figure 4: Visualization results on UCAS-AOD (left column) and CityPersons (right column) datasets.