Research map / Detection and segmentation / 3D detection and driving
DALDet: Depth-Aware Learning Based Object Detection for Autonomous Driving
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DALDet, an efficient one-stage 2D detector that uses stereo-derived depth maps via depth-aware convolution/pooling and a depth-guided loss to improve KITTI 2D detection and output object distance.
Problem
2D object detectors for autonomous driving lack spatial and geometric (depth) cues, and 3D detectors are computationally expensive, so an efficient 2D detector that exploits depth is needed.
Approach
Integrates depth directly into a lightweight one-stage 2D detector via depth-aware convolution and depth-aware average pooling (depth-similarity-weighted operators), and adds a depth-guided loss that penalizes depth-map differences in the non-overlapping regions of predicted and ground-truth boxes.
Key results
- On the KITTI test leaderboard, DALDet reaches 91.01% moderate... wait: moderate Car AP (R40) of 95.88%, ranking 1st among 2D detectors, at 66.7 FPS.
- On KITTI validation, DALDet outperforms YOLOv7 x by 0.24% (Car moderate), 4.01% (Pedestrian) and 21.62% (Cyclist) moderate AP.
- Ablation: full DALDet improves over the no-depth-operator baseline by 3.92% (Car), 6.03% (Pedestrian), 4.93% (Cyclist).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DALDet instead of DSGN++ for 2D Car detection | KITTI test set · KITTI Car moderate | AP (R40): 95.70 -> 95.88 | Helped |
| Use DALDet instead of RRC for 2D Cyclist detection | KITTI test set · KITTI Cyclist moderate | AP (R40): 76.81 -> 79.64 | Helped |
| Use DALDet instead of YOLOv7 x | KITTI validation set · Car moderate | AP (R40): 95.51 -> 95.75 | Helped |
| Use DALDet instead of YOLOv7 x | KITTI validation set · Pedestrian moderate | AP (R40): 72.13 -> 76.14 | Helped |
| Use DALDet instead of YOLOv7 x | KITTI validation set · Cyclist moderate | AP (R40): 60.93 -> 82.55 | Helped |
| Add depth-aware convolution and depth-aware average pooling throughout EDB, FEB, LRB, and FFM (ablation) | DALDet · KITTI validation set | AP2D: 91.83 -> 95.75 (Car); 70.11 -> 76.14 (Pedestrian); 77.62 -> 82.55 (Cyclist) | Helped |
| Add depth-guided loss multiplied with box loss, using α=1, β=0.5, γ=0.05, λ=0.01 (ablation) | DALDet · KITTI validation set | AP2D: 92.88 -> 95.75 (Car); 72.06 -> 76.14 (Pedestrian); 78.72 -> 82.55 (Cyclist) | Helped |
| Use depth-guided loss by multiplication rather than addition (ablation) | DALDet · KITTI validation set | AP2D | Helped |
| Replace IGEV-Stereo depth maps with SGM depth maps (ablation) | DALDet · KITTI validation set | AP2D: 95.75 -> 93.62 (Car); 76.14 -> 72.99 (Pedestrian); 82.55 -> 79.78 (Cyclist) | Hurt |
Surprising findings
- Multiplicative combination of depth-guided loss with box loss works better than additive combination.
Limitations
- Depth is obtained from a separate stereo network (IGEV-Stereo), so the pipeline depends on stereo input and an extra depth model.
- Evaluated only on KITTI Car, Pedestrian and Cyclist classes.
Experimental setup
Datasets: KITTI
Benchmarks: KITTI test leaderboard KITTI validation set
Metrics: AP (R40) FPS
Baselines: YOLOv7 x YOLOv6 m6 YOLOv5 x YOLOX x Sparse R-CNN EfficientDet D4 PP-YOLOE x RRC YOLOv5 x6 DSGN++ EPNet++ PDV
Builds on: Depth-aware convolution (Wang and Neumann 2018) CIoU loss PAFPN IGEV-Stereo
Closest papers by meaning
- IDA-3D: Instance-Depth-Aware 3D Object Detection From Stereo Vision for Autonomous Driving (CVPR 2020 · similarity 0.88)
- CorrBEV: Multi-View 3D Object Detection by Correlation Learning with Multi-modal Prototypes (CVPR 2025 · similarity 0.86)
- PromptDet: A Lightweight 3D Object Detection Framework with LiDAR Prompts (AAAI 2025 · similarity 0.86)
- OpenAD: Open-World Autonomous Driving Benchmark for 3D Object Detection (NeurIPS 2025 · similarity 0.86)
- BEVSpread: Spread Voxel Pooling for Bird's-Eye-View Representation in Vision-based Roadside 3D Object Detection (CVPR 2024 · similarity 0.86)
- SAFDNet: A Simple and Effective Network for Fully Sparse 3D Object Detection (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.