ApolloCar3D: A Large 3D Car Instance Understanding Benchmark for Autonomous Driving

Abstract

Figure 1: An example of our dataset, where (a) is the input color image, (b) illustrates the labeled 2D keypoints, (c) shows the 3D model fitting result with labeled 2D keypoints.

Exploring Deep Learning-Based Architecture, Strategies, Applications and Current Trends in Generic Object Detection: A Comprehensive Review

“…Lu et al [200] proposed novel architecture based on 3D convolutions and RNNs, to generate a centimetre-level localization accuracy in different real-world driving scenarios. 3D car instance understanding and sensor fusion techniques are notable in autonomous driving [201,202]. For further studies, please refer to the recently published survey [203].…”

Simple Baseline for Vehicle Pose Estimation: Experimental Validation

“…As future work, we plan to continue expanding our dataset and studying different keypoints, its viability, and the impact they have on performance as well as explore other architectures, especially those with a bottom-up approach. Additionally, we will go one step further and make the jump from 2D to 3D pose to fully characterise the vehicle structure in a similar way to the one used by [4]- [7], [55].…”

“…In the same way as with stacked hourglass networks, other human pose estimation methods have been used to localise vehicle keypoints. In [7] Song et al used Convolutional Pose Machines (CPMs) [47] as its vehicle keypoint detector. Another interesting approach is the one from Nibali et al in [48].…”

“…Among these works, we can distinguish two groups. On the one hand, those who make use of keypoints [4]- [7], either because they estimate the 2D pose or because they use them as an intermediate step to obtain the 3D pose. On the other hand, those who do not rely on keypoints.…”

“…Besides the aforementioned structural support to improve fine-grained classification, they can be used to enhance instance segmentation like in [23]. Vehicle keypoints can also be used as anchors to retrieve 3D keypoints and structure through the use of CAD models and camera parameters as in [4]- [7]. They can help to improve traffic surveillance applications [24], including vehicle reidentification [25], when a vehicle has to be identified on multiple images from different cameras, and license plate recognition is not possible (usually because the camera resolution is very poor or the license plate is not visible).…”