Abstract
Monocular 3D object detection aims to locate objects in different scenes with just a single image. Due to the absence of depth information, several monocular 3D detection techniques have emerged that rely on auxiliary depth maps from the depth estimation task. There are multiple approaches to understanding the representation of depth maps, including treating them as pseudo-LiDAR point clouds, leveraging implicit end-to-end learning of depth information, or considering them as an image input. However, these methods have certain drawbacks, such as their reliance on the accuracy of estimated depth maps and suboptimal utilization of depth maps due to their image-based nature. In this paper, we propose a framework named the Adaptive Distance Interval Separation Network (ADISN) that adopts a novel perspective on understanding depth maps, as a form that lies between LiDAR and images. We utilize an adaptive separation approach that partitions the depth map into various subgraphs based on distance and treats each of these subgraphs as an individual image for feature extraction. After adaptive separations, each subgraph solely contains pixels within a learned interval range. If there is a truncated object within this range, an evident curved edge will appear, which we can leverage for texture extraction using CNNs to obtain rich depth information in pixels. Meanwhile, to mitigate the inaccuracy of depth estimation, we designed an uncertainty module. To take advantage of both images and depth maps, we use different branches to learn localization detection tasks and appearance tasks separately.
Introduction
Monocular 3D object detection—identifying the position, size, and orientation of objects from a single 2D image—is a critical challenge for low-cost autonomous systems. Unlike expensive LiDAR sensors that provide direct 3D spatial data, a single camera lacks inherent depth information. To bridge this gap, researchers often utilize estimated depth maps as auxiliary inputs.
However, existing methods typically treat these maps either as standard 2D images, which ignores their physical spatial meaning, or as "pseudo-LiDAR" point clouds, which are highly sensitive to depth estimation errors. The Adaptive Distance Interval Separation Network (ADISN) introduces a middle ground: it processes depth data in a way that respects its physical, distance-based nature while leveraging the feature extraction power of Convolutional Neural Networks (CNNs).
Methodology
Adaptive Distance Interval Separation (ADIS)
The core innovation of ADISN is the ADIS module. Instead of feeding a raw, continuous depth map into a CNN, the network learns to slice the depth map into several "sub-depth maps" (SDs) based on adaptive distance intervals. For instance, one slice might contain pixels representing objects 5–10 meters away, while another covers 10–20 meters.
This thresholding creates sharp artificial edges where objects are "cut" by the interval boundaries. Since CNNs are naturally optimized for edge and texture detection, these manufactured boundaries provide strong, high-contrast cues for object localization that are often missing in smooth, raw depth maps.
Uncertainty-Aware Localization
Monocular depth estimation is notoriously noisy, particularly for distant objects or complex backgrounds like the sky. To mitigate this, ADISN incorporates an uncertainty module that learns a confidence map for the estimated depth. This allows the model to dynamically "distrust" unreliable depth regions, shifting its reliance toward the original RGB image in those areas to maintain detection robustness.
Feature Decoupling and Architecture
The ADISN architecture utilizes separate "heads" to handle distinct aspects of 3D detection. An appearance branch focuses on category classification and visual features by fusing RGB and depth data. Simultaneously, a localization branch utilizes the ADIS features and the uncertainty map to pinpoint the object's precise 3D coordinates $(x, y, z)$.
Experimental Results
The model was rigorously evaluated on the KITTI monocular 3D object detection benchmark, which includes 7,481 training samples across Car, Pedestrian, and Cyclist categories. ADISN achieved state-of-the-art performance for the "Car" category, ranking 1st among monocular methods in Easy, Moderate, and Hard difficulty levels for 3D Average Precision (AP 3D). Notably, it outperformed the established M3D-RPN baseline by 3.11% on Easy tasks and 2.43% on Moderate tasks. Qualitative results indicate that the model is particularly effective at detecting distant or partially occluded vehicles that traditional models often overlook.
Conclusion
ADISN demonstrates that depth maps are most effective when treated as a hybrid representation between 2D images and 3D point clouds. By "manufacturing" edges through distance-based separation, the framework allows standard vision architectures to extract precise localization features. While the system remains somewhat dependent on the quality of the initial depth estimator, the inclusion of an uncertainty module provides a necessary layer of robustness for real-world autonomous driving scenarios.
Summary updated