Evaluating foundational depth models for zero-shot 3D box lifting from 2D detections

Official URL: https://dx.doi.org/10.1109/ICCE67443.2026.11449883

Monocular 3D object detection offers a low-cost alternative to LiDAR-based perception, but remains fundamentally limited by scale ambiguity and the absence of geometric depth cues in single images. Although recent deep learning approaches have advanced performance, they typically rely on large-scale annotated datasets and complex, fully supervised pipelines. In this work, we introduce a benchmarking framework designed to systematically evaluate zero-shot 3D object lifting from 2D detections using pre-trained monocular depth estimation models. Rather than proposing a new detector, our tool provides a modular and training-free platform that combines off-the-shelf 2D detectors, camera calibration, and existing depth models to project detections into 3D space. Our method does not require labeled 2D or 3D data, enabling flexible deployment across datasets and domains. The modular design allows plug-and-play evaluation of both 2D detectors and monocular depth models, facilitating controlled benchmarking without retraining the entire model. We evaluate our framework on the KITTI benchmark and show that, despite not using any learning or fine-tuning, it achieves competitive 3D detection results, particularly when strong foundation models are used. Our approach not only provides a practical tool for rapid 3D perception in label-scarce scenarios, but also serves as a baseline to analyze the effectiveness of different depth models for object-level spatial reasoning.