3D object detection is gradually becoming more important in our lives. An easy example would be the recent production of autonomous vehicles: they scan their surroundings and understand what they are. Since the technology of said cars is yet unripe, there are many gaps to close for scientists. For example, current high-performing software uses the center of an object for detection, so they are both unable to relate spaced out information to one another and confused when the middle of the object is empty. Hierarchal Encored-Decoder Network (HEDNet) tries to combat this problem by adding encoder-decoder blocks to collect and process details far apart, usually, for large objects. By applying the principle, HEDNet achieved unparallel accuracy and efficiency.
HEDNet Vs. Older Software
Currently, the state-of-the-art method relies on voxel-based detection. The process involves breaking down big data into smaller chunks and distributing them to small kernels with submanifold sparse residual (SSR) blocks. Even though this way of handling and analyzing data is effective, kernels not communicating with each other limit the range of features that can be prescribed to each object. Researchers have tried to overcome this issue by using large kernels; however, this only provided restricted improvements and a huge increase in computational costs. HEDNet on the other hand, uses sparse encoder-decoder (SED) and dense encoder-decoder (DED) blocks instead of SSR blocks. SED blocks are designed to reduce the distance between features of a certain object through down-sampling, only receiving the important parts. They then up-sample the resultant map to match the resolution of the first one, but in doing so, the blocks only up-sample the valid features and eliminate excess information. DED blocks are built similarly but have a different purpose. Since current detection systems are reliant on object centers, DED blocks help expand features towards the center to increase the accuracy of detection. HEDNet and other competitor software were tested using the datasets nuScenes and Waymo Open. The addition of SED and DED blocks put a clear gap between HEDNet and other software. While it may have shown supreme accuracy in outdoor scenarios, HEDNet still must be worked on to be used indoors.
Usage
HEDNet specializes in self-driving cars and their environmental detection, but there are more areas where this technology could be used. The industrial machinery industry has slowly begun using similar techniques for both safety and efficiency purposes. The agriculture industry uses it for monitoring plants. Security companies may utilize it for video surveillance. Finally, several jobs might benefit from autonomously 3D scanning objects.