Improving object detection for truck-related classes by removing label inconsistencies

Accurate detection of a wide variety of objects is critically important for object detection in autonomous driving. However, many autonomous driving datasets are dominated by cars and pedestrians, while large vehicles such as trucks are both underrepresented and more diverse in size and appearance than the majority classes. One challenge is that trucks generally consist of multiple components (tractors and trailers), which can overlap with each other and may also be seen alone. Additionally, there is no consensus on the optimal class definitions for such objects: some datasets have separate classes for truck components or specific types of truck, and it is not uncommon for subordinate level (individual objects) and superior level (group of objects) definitions to be mixed in the same dataset. We believe that this inconsistency in labeling diminishes perception performance by introducing semantic confusion for object components with multiple valid class definitions. In this work, we test multiple labeling approaches for truck objects using a multiclass LiDAR-only object detection network. We consider labeling schemes including: both subordinate and superior class descriptions as a baseline, subordinate and superior class labels combined into one class, elimination of the superior class description (truck), and elimination of the subordinate class descriptions (truck tractor and truck trailer). We find that using only one descriptive level to categorize related objects is important: using only the subordinate classes or superior class yielded better results than mixing the two. Furthermore, we find that using the subordinate classes “Truck Front” and “Truck Trailer” yields the best overall performance and better captures trucks making wide turns, which is not well represented in many datasets.