A forensic evaluation framework for egocentric foundation models: an EFM3D case study for crime scene reconstruction
Organizations: School of Computing, Engineering & Technology, Robert Gordon University, Aberdeen, United Kingdom
Abstract
Foundation models increasingly combine geometry, semantics, and spatial context for three-dimensional scene understanding, but their suitability for forensic crime scene reconstruction remains unclear. This study presents a structured forensic evaluation framework for egocentric foundation models and demonstrates it using the Egocentric Foundation Model for 3D (EFM3D). Four simulated indoor crime scenes were recorded with Meta Project Aria Gen~1 at Robert Gordon University's East Lodge House forensic training facility. The framework evaluates room-scale reconstruction, semantic scene understanding, computational performance, and forensic suitability. Across the four scenes, EFM3D generated meshes averaging vertices and faces, with 39,529 recorded semantic predictions spanning 21--24 categories at a mean confidence of 0.40. Because reference meshes and expert-labelled 3D ground truth were unavailable, these statistics are treated as descriptive outputs rather than measures of geometric or detection accuracy. Final inference runs processed snippets per scene at samples/s and required minutes per recording. A seven-criterion capability profile indicated stronger support for room geometry, furniture localisation, scene context, and spatial relationships than for fine-scale forensic evidence. The results position EFM3D as a complementary tool for contextual documentation and scene review rather than a replacement for high-fidelity forensic capture. The framework provides a reproducible basis for evaluating future egocentric foundation models in specialised forensic settings.
Figures & tables
| Characteristic | ADT [ 15 ] | AEA [ 11 ] | Forensic training facility |
|---|---|---|---|
| Environment | Furnished apartment and office | Everyday activities in real homes | Four staged indoor crime-scene rooms |
| Sequences | 200 | 143 | 4 |
| Volume | Sequence count reported | 7.3 h; 1M frames | 563.7 s; 16,921 RGB frames |
| Object ground truth | 398 instances with 6DoF poses and segmentation | None published | No manual ground truth; EFM3D outputs only |
| Semantic scope | Everyday household/office objects | No released semantic labels | 21–24 predicted everyday categories; no forensic classes |
| Scene intent | Natural domestic/office activity | Natural daily activity | Disturbed layouts and forensic walkthroughs |
| Metric | Room 1 | Room 2 | Room 3 | Room 4 | Mean |
|---|---|---|---|---|---|
| Duration (s) | 112.61 | 157.54 | 153.39 | 140.15 | 140.92 |
| VRS size (GB) | 1.314 | 1.962 | 1.762 | 1.604 | 1.661 |
| RGB frames | 3,380 | 4,729 | 4,605 | 4,207 | 4,230 |
| RGB rate (fps) | 30.02 | 30.02 | 30.02 | 30.02 | 30.02 |
| Stream | Resolution | Rate | Format |
|---|---|---|---|
| RGB camera | px | 30 fps | JPEG |
| SLAM cameras ( ) | px | 30 fps | JPEG |
| Eye cameras ( ) | px | 60 fps | JPEG |
| IMU 1 / IMU 2 | — | 1000 / 800 Hz | — |
| Magnetometer / barometer | — | 10 / 50 Hz | — |
| Microphones ( ) | — | 48 kHz | — |
| Component | Specification |
|---|---|
| Wearable sensor | Meta Project Aria Gen 1 |
| Recording profile | Profile 28 |
| Operating system | Ubuntu 24.04 LTS |
| Development environment | Google Colaboratory |
| Programming language | Python 3.12 |
| GPU | NVIDIA A100-SXM4-40GB |
| Metric | R1 | R2 | R3 | R4 | Mean SD |
|---|---|---|---|---|---|
| Duration (s) | 112.61 | 157.54 | 153.39 | 140.15 | 140.92 20.30 |
| RGB frames | 3,380 | 4,729 | 4,605 | 4,207 | 4,230 608 |
| RGB rate (fps) | 30.02 | 30.02 | 30.02 | 30.02 | 30.02 0.00 |
| Mesh vertices | 10,485 | 12,203 | 16,235 | 11,750 | 12,668 2,430 |
| Mesh faces | 20,590 | 23,811 | 31,904 | 23,045 | 24,838 4,903 |
| Metric | Value |
|---|---|
| Total recorded predictions | 39,529 |
| Mean predictions per scene | 9,882 |
| Categories per scene | 21–24 |
| Mean confidence | 0.400 |
| Mean predictions per RGB frame | 2.34 |
| Mean simultaneous predictions | 7.14 |
| Metric | R1 | R2 | R3 | R4 | Mean SD |
|---|---|---|---|---|---|
| Snippets processed | 1,105 | 1,554 | 1,513 | 1,380 | 1,388 203 |
| Throughput (samples/s) | 0.70 | 0.69 | 0.69 | 0.67 | 0.69 0.01 |
| Inference time (s) | 2580.41 | 3408.28 | 3257.84 | 2938.72 | 3046.31 367.15 |
| Inference time (min) | 43.01 | 56.80 | 54.30 | 48.98 | 50.77 6.12 |
| VRS size (GB) | 1.314 | 1.962 | 1.762 | 1.604 | 1.661 0.278 |
| Criterion | Capability | Basis in observed outputs |
|---|---|---|
| Room geometry reconstruction | Supported | Recognisable room-scale meshes in all four scenes |
| Furniture localisation | Supported | Common furniture predictions present across scenes |
| Scene context understanding | Partially supported | Broad indoor semantic context represented, with limitations |
| Spatial relationships | Supported | Room-scale geometry and object locations jointly represented |
| Small-evidence representation | Limited | Qualitative inspection indicated weak/inconsistent representation |
| Bloodstain recognition | Not supported | Forensic-specific class absent from model vocabulary |
3D Foundation Models
3D foundation models learn reusable geometric and semantic representations for reconstruction, recognition, generation, and spatial reasoning.
See all 71 papers