Two JD Retail Papers Accepted at ICRA 2021: Long‑tailed Facial Expression Recognition and Vanishing‑Point‑Aided LiDAR‑Visual‑Inertial Estimator

ICRA (International Conference on Robotics and Automation) is a top‑tier annual academic conference in the robotics field. Recently, JD Retail’s Technology and Data Center – Shared Technology Department announced that two of its papers were selected for presentation.

Paper 1: Deep Balanced Learning for Long‑tailed Facial Expressions Recognition

The authors address the data‑imbalance problem in facial expression datasets, where common emotions such as joy dominate while others are scarce. They propose a structure‑invariant resampling method combined with a two‑stage training strategy that prunes and fixes a portion of network weights, achieving balanced training without extra parameters.

Experimental results on the AffectNet dataset show that the proposed method outperforms state‑of‑the‑art techniques and can be applied to any backbone network, offering a practical solution for real‑world deployment.

Paper 2: Vanishing Point‑Aided LiDAR‑Visual‑Inertial Estimator

This work targets robust and accurate state estimation for autonomous driving, drones, and AR/VR. It introduces a three‑module system: (1) an IMU‑assisted vanishing‑point detection module using 1‑line RANSAC to establish correspondences across frames; (2) a voxel‑map‑based feature‑depth association module that assigns depth to visual features; (3) a visual‑inertial lag‑correction smoothing module that jointly minimizes error terms.

Experiments in indoor and outdoor scenarios demonstrate that the proposed estimator surpasses current leading visual‑inertial odometry and LiDAR‑visual estimation methods.

The Shared Technology Department will continue to deepen research in AR/VR and computer vision, striving for innovative breakthroughs and practical applications.