| { |
| "field_id": "F003", |
| "field": "3d object detection", |
| "directions": [ |
| { |
| "direction_id": "G0", |
| "name": "3D detection from camera geometry", |
| "scope": "Infer the 3D position and shape needed for detection from camera images, including depth and geometry inference from a single image, shared 3D representations built from multiple views, and depth estimation through explicit binocular or cross-temporal image correspondence." |
| }, |
| { |
| "direction_id": "G1", |
| "name": "Detection evidence from range measurements", |
| "scope": "Make spatial evidence from point-cloud or radar measurements sufficient for 3D object detection by learning geometric features from observed points, compensating for incomplete evidence caused by sparsity, distance, or occlusion, and addressing radar-specific sparsity, noise, and encoding." |
| }, |
| { |
| "direction_id": "G2", |
| "name": "Computationally feasible 3D detection", |
| "scope": "Reduce the computational and memory costs of detection representations, attention, or spatial processing while examining their detection effectiveness." |
| }, |
| { |
| "direction_id": "G3", |
| "name": "Cross-sensor evidence fusion for detection", |
| "scope": "Associate, align, and combine complementary evidence actually available from different sensors during detection to improve 3D localization and recognition." |
| }, |
| { |
| "direction_id": "G4", |
| "name": "Temporal evidence for 3D detections", |
| "scope": "Align, retain, or propagate evidence from past frames, scans, and object motion to improve detections, rather than making identity maintenance the primary output." |
| }, |
| { |
| "direction_id": "G5", |
| "name": "Supplementary training evidence for 3D detectors", |
| "scope": "Improve detectors using knowledge or supervision beyond direct training on 3D boxes, including teacher-to-student transfer of detection knowledge, detection-oriented representation pre-training, and training with weak constraints and pseudo-labels when 3D box annotations are limited or absent." |
| }, |
| { |
| "direction_id": "G6", |
| "name": "Detection across shifted and degraded conditions", |
| "scope": "Study how trained detectors remain effective when target conditions change: address training-to-deployment distribution differences in scenes, sensor sampling, or object geometry, and address corrupted, missing, adverse, or attacked input observations." |
| }, |
| { |
| "direction_id": "G7", |
| "name": "Assistance with 3D box annotation", |
| "scope": "Reduce the human effort required to create or correct 3D object-box annotations, with usable annotations—rather than a detector trained on a fixed annotation set—as the primary output." |
| } |
| ] |
| } |
|
|
