SE-SSD: Self-Ensembling Single-Stage Object Detector From Point Cloud

Abstract: We present Self-Ensembling Single-Stage object Detector (SE-SSD) for accurate and efficient 3D object detection in outdoor point clouds. Our key focus is on exploiting both soft and hard targets with our formulated constraints to jointly optimize the model, without introducing extra computation in the inference. Specifically, SE-SSD contains a pair of teacher and student SSDs, in which we design an effective IoU-based matching strategy to filter soft targets from the teacher and formulate a consistency loss to align student predictions with them. Also, to maximize the distilled knowledge for ensembling the teacher, we design a new augmentation scheme to produce shape-aware augmented samples to train the student, aiming to encourage it to infer complete object shapes. Lastly, to better exploit hard targets, we design an ODIoU loss to supervise the student with constraints on the predicted box centers and orientations. Our SE-SSD attains top performance compared with all prior published works. Also, it attains top precisions for car detection in the KITTI benchmark (ranked 1st and 2nd on the BEV and 3D leaderboards, respectively) with an ultra-high inference speed. The code is available at https://github.com/Vegeta2020/SE-SSD.

• The paper presents a dual-network (teacher-student) architecture that utilizes self-ensembling to improve robustness in 3D object detection.
• The paper employs an IoU-based matching strategy and shape-aware augmentation to generate reliable soft targets and refine geometric predictions.
• The paper demonstrates its effectiveness on the KITTI benchmark, achieving 1st in BEV and 2nd in 3D detection for cars with minimal inference overhead.

Overview of Self-Ensembling Single-Stage Object Detector (SE-SSD)

The paper introduces the Self-Ensembling Single-Stage Object Detector (SE-SSD), a novel approach targeting efficient and accurate 3D object detection in outdoor point clouds. This method is designed to leverage self-ensembling principles specifically tailored for 3D object detection tasks, a domain increasingly pertinent due to advancements in autonomous systems and robotics.

Methodology

SE-SSD's architecture is centered around a dual-network framework comprising a teacher and a student SSD. The interaction between these networks is critical for the model's success. Key elements include:

• IoU-based Matching Strategy: The teacher network generates soft target proposals that serve as reference points for the student network. Through an Intersection over Union (IoU) criterion, the system filters these proposals to ensure high-quality guidance.
• Consistency Loss: A unique loss formulation maintains coherence between the student’s output and the filtered soft targets from the teacher. This aids the student network in learning more robust representations by aligning its predictions with reliable teacher suggestions.
• Shape-aware Augmentation: To enhance the student network's understanding of object geometries, a novel augmentation scheme is implemented. This technique focuses on producing shape-consistent augmented samples, thus enabling the student network to infer comprehensive object delineations even from partial observations.
• ODIoU Loss: The orientation and location of object predictions are explicitly refined through the Optimized Dense IoU (ODIoU) loss. This suppresses errors in bounding box dimensions and orientations by enforcing constraints directly on predicted box centers and orientations.

Results

The SE-SSD model achieves exemplary performance on the KITTI benchmark, particularly in car detection tasks. Its results are substantial enough to achieve top rankings—1st in BEV and 2nd in 3D leaderboards—demonstrating its competitive edge over extant methodologies. Importantly, this is achieved with minimal inference overhead, ensuring the model's applicability in real-time systems.

Implications and Future Work

The SE-SSD framework's success underscores the efficacy of self-ensembling techniques in object detection within point cloud environments. The proposed IoU-based strategy, coupled with shape-aware augmentation, presents a compelling paradigm for self-supervised learning scenarios.

Further developments could explore extending these principles to other domains and object types, potentially incorporating different sensory inputs or more complex environmental conditions. Additionally, future research might refine augmentation strategies to adapt dynamically based on scene characteristics or object occlusion patterns.

By releasing the codebase publicly, the authors facilitate further exploration and validation, encouraging extensions and adaptations. The ongoing evolution in 3D detection and this paper's contributions highlight the trajectory towards more sophisticated perception systems in both academic settings and industry applications.