Research map / Architectures and efficiency / CNN design
EDENet: Echo Direction Encoding Network for Place Recognition Based on Ground Penetrating Radar
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 EDENet, a compact GPR sequence-descriptor network built from learnable Gabor filters, direction-aware attention, and multi-scale aggregation for large-scale ground-penetrating-radar place recognition.
Problem
Place recognition with ground penetrating radar (GPR) in large-scale maps is hard due to sparse underground features and variable dielectric constants that distort echo features.
Approach
Encodes the directional geometry of GPR echoes using a bank of learnable Gabor filters, recalibrates filter responses with a direction-aware attention module (with skip connection), adds a shift-invariant max-pooling unit, and aggregates multi-scale EDE blocks into a compact L2-normalized descriptor.
Key results
- On GROUNDED Sunny/Rainy, EDENet reaches recall@1/5/10 of 0.71/0.81/0.85 versus 0.54/0.60/0.67 for LGPRNet and 0.50/0.61/0.69 for CMUNet.
- On CMU-GPR Nsh_b, EDENet reaches 0.81/0.86/0.89 recall@1/5/10, versus 0.57/0.69/0.76 for MWSNet.
- EDENet has 5.0x10^8 parameters, the smallest among compared methods, and runs at 188 Hz on 100-frame sequences (0.005s), versus a 126 Hz GPR acquisition rate.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EDENet instead of the strongest listed baseline | GPR place recognition · GROUNDED Sunny/Rainy | recall@1/5/10: 0.54/0.60/0.67 -> 0.71/0.81/0.85 | Helped |
| Use EDENet instead of the strongest listed baseline | GPR place recognition · GROUNDED Sunny/Snowy | recall@1/5/10: 0.47/0.56/0.62 -> 0.57/0.66/0.70 | Helped |
| Use EDENet instead of MWSNet | GPR place recognition · CMU-GPR Nsh_b | recall@1/5/10: 0.57/0.69/0.76 -> 0.81/0.86/0.89 | Helped |
| Use EDENet instead of MWSNet | GPR place recognition · CMU-GPR Gate_g | recall@1/5/10: 0.81/0.94/0.95 -> 0.95/0.99/0.99 | Helped |
| Add learnable Gabor filters instead of standard convolution (ablation) | EDENet · GROUNDED | recall@1/5: 0.63/0.68 -> 0.71/0.76 | Helped |
| Add learnable Gabor filters instead of standard convolution (ablation) | EDENet · CMU-GPR | recall@1/5: 0.71/0.78 -> 0.75/0.82 | Helped |
| Add the Direction Aware Attention module (ablation) | EDENet · GROUNDED | recall@1/5: 0.62/0.70 -> 0.71/0.76 | Helped |
| Add the Direction Aware Attention module (ablation) | EDENet · CMU-GPR | recall@1/5: 0.74/0.80 -> 0.75/0.82 | Helped |
| Use EDE-35, EDE-23, EDE-11, and EDE-5 scales instead of removing EDE-23 (ablation) | EDENet with k=64 · GROUNDED | recall@5: 74.6 -> 76.0 | Helped |
| Use EDE-35, EDE-23, EDE-11, and EDE-5 scales instead of removing EDE-5 (ablation) | EDENet with k=64 · GROUNDED | recall@5: 73.2 -> 76.0 | Helped |
| Use EDE-35, EDE-23, EDE-11, and EDE-5 scales instead of removing EDE-35 (ablation) | EDENet with k=64 · GROUNDED | recall@5: 68.6 -> 76.0 | Helped |
| Use EDE-35, EDE-23, EDE-11, and EDE-5 scales instead of retaining only EDE-35 and EDE-5 (ablation) | EDENet with k=64 · GROUNDED | recall@5: 72.6 -> 76.0 | Helped |
Surprising findings
- Camera and LiDAR sequence PR methods (SeqVLAD, SeqOT) overfit on GPR data, with SeqVLAD performing poorly on CMU-GPR (0.23/0.27/0.29 recall@1/5/10 on Nsh_b).
Limitations
- Parameters are concentrated in the final feature aggregation MLP layer, so the model is not as lightweight as claimed.
- Learnable Gabor filters are not orthogonal and can overlap, causing redundancy and overfitting as scales or k increase.
- Evaluated only on two public GPR datasets (GROUNDED and CMU-GPR).
Experimental setup
Datasets: GROUNDED CMU-GPR
Benchmarks: GROUNDED CMU-GPR
Metrics: recall@1 recall@5 recall@10 parameters model size runtime
Baselines: 3D ResNet18 3D ResNet50 3D ResNet101 CMUNet DEC MWSNet SeqVLAD SeqOT LGPRNet
Builds on: SeqVLAD SeqOT NetVLAD Gabor filter
Closest papers by meaning
- TransLoc4D: Transformer-based 4D Radar Place Recognition (CVPR 2024 · similarity 0.81)
- Multi-Modal Aerial-Ground Cross-View Place Recognition with Neural ODEs (CVPR 2025 · similarity 0.80)
- Deep Homography Estimation for Visual Place Recognition (AAAI 2024 · similarity 0.80)
- GeoRanker: Distance-Aware Ranking for Worldwide Image Geolocalization (NeurIPS 2025 · similarity 0.79)
- ForestLPR: LiDAR Place Recognition in Forests Attentioning Multiple BEV Density Images (CVPR 2025 · similarity 0.79)
- Geographic Location Encoding with Spherical Harmonics and Sinusoidal Representation Networks (ICLR 2024 · similarity 0.79)
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.