Enhancing Precision of object detectors: bridging classification and localization gaps for 2D and 3D models

Abstract

Artificial Intelligence (AI) has revolutionized and accelerated significant advancements in various fields such as healthcare, finance, education, agriculture and the development of autonomous vehicles. We are rapidly approaching Level 5 Autonomy due to recent developments in autonomous technology, including self-driving cars, robot navigation, smart traffic monitoring systems, and dynamic routing. This success has been made possible due to Deep Learning technologies and advanced Computer Vision (CV) algorithms. With the help of perception sensors such as Camera, LiDAR and RADAR, CV algorithms enable a self-driving vehicle to interact with the environment and make intelligent decisions. Object detection lays the foundations for various applications, such as collision and obstacle avoidance, lane detection, pedestrian and vehicular safety, and object tracking. Object detection has two significant components: image classification and object localization. In recent years, enhancing the performance of 2D and 3D object detectors has spiked interest in the research community. This research aims to resolve the drawbacks associated with localization loss estimation of 2D and 3D object detectors by addressing the bounding box regression problem, addressing the class imbalance issue affecting the confidence loss estimation, and finally proposing a dynamic cross-model 3D hybrid object detector with enhanced localization and confidence loss estimation. This research aims to address challenges in object detectors through four key contributions. In the first part, we aim to address the problems associated with the image classification component of 2D object detectors. Class imbalance is a common problem associated with supervised training. Common causes are noisy data, a scene with a tiny object surrounded by background pixels, or a dense scene with too many objects. These scenarios can produce many negative samples compared to positive ones, affecting the network learning and reducing the overall performance. We examined these drawbacks and proposed an Enhanced Hard Negative Mining (EHNM) approach, which utilizes anchor boxes with 20% to 50% overlap and positive and negative samples to boost performance. The efficiency of the proposed EHNM was evaluated using Single Shot Multibox Detector (SSD) architecture on the PASCAL VOC dataset, indicating that the detection accuracy of tiny objects increased by 3.9% and 4% and the overall accuracy improved by 0.9%. To address localization loss, our second approach investigates drawbacks associated with existing bounding box regression problems, such as poor convergence and incorrect regression. We analyzed various cases, such as when objects are inclusive of one another, two objects with the same centres, two objects with the same centres and similar aspect ratios. During our analysis, we observed existing intersections over Union (IoU) loss and its variant’s failure to address them. We proposed two new loss functions, Improved Intersection Over Union (IIoU) and Balanced Intersection Over Union (BIoU), to enhance performance and minimize computational efforts. Two variants of the YOLOv5 model, YOLOv5n6 and YOLOv5s, were utilized to demonstrate the superior performance of IIoU on PASCAL VOC and CGMU datasets. With help of ROS and NVIDIA’s devices, inference speed was observed in real-time. Extensive experiments were performed to evaluate the performance of BIoU on object detectors. The evaluation results indicated MASK_RCNN network trained on the COCO dataset, YOLOv5n6 network trained on SKU-110K and YOLOv5x trained on the custom e-scooter dataset demonstrated 3.70% increase on small objects, 6.20% on 55% overlap and 9.03% on 80% overlap. In the earlier parts, we primarily focused on 2D object detectors. Owing to its success, we extended the scope of our research to 3D object detectors in the later parts. The third portion of our research aims to solve bounding box problems associated with 3D rotated objects. Existing axis-aligned loss functions suffer a performance gap if the objects are rotated. We enhanced the earlier proposed IIoU loss by considering two additional parameters: the objects’ Z-axis and rotation angle. These two parameters aid in localizing the object in 3D space. Evaluation was performed on LiDAR and Fusion methods on 3D KITTI and nuScenes datasets. Once we addressed the drawbacks associated with confidence and localization loss, we further explored ways to increase the performance of cross-model 3D object detectors. We discovered from previous studies that perception sensors are volatile to harsh environmental conditions, sunlight, and blurry motion. In the final portion of our research, we propose a hybrid 3D cross-model detection network (MAEGNN) equipped with MaskedAuto Encoders (MAE) and Graph Neural Networks (GNN) along with earlier proposed IIoU and ENHM. The performance evaluation on MAEGNN on the KITTI validation dataset and KITTI test set yielded a detection accuracy of 69.15%, 63.99%, 58.46% and 40.85%, 37.37% on 3D pedestrians with overlap of 50%. This developed hybrid detector overcomes the challenges of localization error and confidence estimation and outperforms many state-of-art 3D object detectors for autonomous platforms.