Tackling the Issues of 3D Object Processing in Autonomous Vehicles

    Autonomous driving vehicles rely on a variety of sources to be able to gather as much information as possible about its surroundings so that it can make decisions quickly. These vehicles especially depend on depth sensors (RGB-D cameras), light detection and ranging (LiDAR) and radio detection and ranging (radar) so that they are able to fully grasp the surroundings. Even with all of this special equipment used, the perception systems of autonomous vehicles are still making errors which could be detrimental in comparison to a human driving the vehicle, and should therefore be properly addressed. 

    In the article: “Beyond the Contour: How 3D Cues Enhance Object Recognition in Humans and Neural Networks”, Cutler and colleagues (2025) analyzed the differences between how humans and artificial neural networks (ANNs) are able to determine what an object is based on different cues. Some of these cues included the outline, textures, shapes, shades and shadows. The researchers were able to determine that although humans are more biased to shapes when detecting objects, they are still able to rely upon the texture cues. They also concluded that ANNs are more texture biased when trying to determine what an object is. Furthermore, when they added shades and shadows to enhance the 3D objects, the researchers found that these were advantageous features to be able to recognize objects for both humans and ANNs (Cutler et al., 2025). Although artificial neural networks are inherently more texture biased, they can be trained enough with other object cues to expand their networks and eliminate these biases. This study demonstrates the importance of continuous training with artificial models so that they are able to perform like humans and eventually be even better than humans. 

        The findings of Cutler et al. (2025) can provide the basis to explain how the networks in autonomous vehicles are gathering information to be properly equipped to fully function without humans. These processing networks still rely upon 2D object detection which is primarily texture biased as previous research has stated. Through significant amounts of training and growth of deep learning (DL), Alaba and Ball (2023) also found that the networks have improved in their performance which demonstrates that it is advantageous to continuously train their artificial models. In the review article: “Deep Learning-Based Image 3-D Object Detection for Autonomous Driving: Review”, Alaba and Ball (2023) also explained that these autonomous vehicles initially used a two-stage object processing system in which they first analyze the object in its 2D form and then use different techniques to examine its 3D form. It was then proposed that this two-stage process be reduced to a single stage so that these processing networks are more efficient, especially with the new processing network called YOLO (you only look once). Although it became more efficient, this system was less accurate, especially because it was processing a whole scene which can include a variety of objects of many different shapes and sizes and would therefore have to classify everything at once. In more recent times, the YOLO processing system has been further worked on to improve its accuracy to the same level as the two-stage process that was previously used. Overall, this demonstrates that work with artificial models and networks must be ongoing to continue to improve on its accuracy and efficiency, especially in real world contexts like driving. 

References: 

Alaba, S. Y., & Ball, J. E. (2023). Deep Learning-Based Image 3-D Object Detection for Autonomous Driving: Review. IEEE Sensors Journal, 23(4), 3378–3394. https://doi.org/10.1109/jsen.2023.3235830

Cutler, M., Baumel, L. D., Tocco, J., Friebel, W., Thiruvathukal, G. K., & Baker, N. (2025). Beyond the contour: How 3D cues enhance recognitions in humans and neural networks. Journal of Vision. (Pre-published draft)