Search

Accurately detect vehicles or pedestrians from 3D point clouds (3D object detection) is a fast developing research topic in autonomous driving and other domains. The fundamental component for feature extraction in 3D object detection is Set Abstraction (SA), which can downsample points while aggregating points to extract features. However, the current SA ignores the geometric and semantic properties of point clouds and may miss to detect remote small objects. In this paper, FocusSA is proposed, which consists two modules for enhancing useful feature extraction in the SA layer to improve 3D object detection accuracy. At first, Focused FPS (FocFPS) is proposed to evaluate the foreground and boundary scores of the points and reweighs the Furthest Point Sampling (FPS) using the evaluated scores to retain more contextual points in downsampling. Then a Geometry-aware Feature Extraction (GeoFE) module is proposed to add geometric information to enrich the awareness of geometric structure in feature aggregation. To evaluate the performances of the proposed methods, we conduct extensive experiments on three difficulty levels of Car class in KITTI dataset. The experimental results show that on "moderate" instances, our results outperform the state-of-the-art method by 1.08%. Moreover, FocusSA is easy to be plugged in popular architectures. [ABSTRACT FROM AUTHOR]

Autonomous driving and the machine learning (ML) models developed to achieve it have grown rapidly in importance and complexity respectively, and with them has also grown their carbon footprint due to long training times. Given the importance of climate change, it should be necessary to include the CO 2 impact of ML models explicitly to encourage competition on more than just model quality. This work presents the implementation of two different Convolutional Neural Network (CNN) training approaches, used for autonomous driving by behavioral cloning in a simulated environment and compares their impact on the CO 2 footprint. Using a cloud execution environment and driving data, previously obtained by applying an end-to-end deep learning technique, the first implemented training approach is a classical approach that uses an image generator that carries out the pre-processing and data augmentation during training. The contribution proposed in this paper is a second approach that improves the training by decreasing the training time, carrying out the data augmentation and pre-processing tasks before training the model, storing the result in RAM and then starting the training. The new approach to training presented in this article finishes the training approximately 38 times faster and reduces the carbon footprint impact by approximately 96%. In absolute values, this is a reduction from an average value of approximately 0.1643 CO 2 e q (kg) to 0.007 CO 2 e q (kg). To estimate the impact of CO 2 , the hardware used for the project, the training time and the cloud service provider were all taken into account. [ABSTRACT FROM AUTHOR]