cs.CVMar 19, 2026
SaveFoundations and Architectures of Artificial Intelligence for Motor Insurance
Abstract
This handbook presents a systematic treatment of the foundations and architectures of artificial intelligence for motor insurance, grounded in large-scale real-world deployment. It formalizes a vertically integrated AI paradigm that unifies perception, multimodal reasoning, and production infrastructure into a cohesive intelligence stack for automotive risk assessment and claims processing. At its core, the handbook develops domain-adapted transformer architectures for structured visual understanding, relational vehicle representation learning, and multimodal document intelligence, enabling end-to-end automation of vehicle damage analysis, claims evaluation, and underwriting workflows. These components are composed into a scalable pipeline operating under practical constraints observed in nationwide motor insurance systems in Thailand. Beyond model design, the handbook emphasizes the co-evolution of learning algorithms and MLOps practices, establishing a principled framework for translating modern artificial intelligence into reliable, production-grade systems in high-stakes industrial environments.
Figures & tables
Figure 1.1: MARSAIL Artificial Intelligence Laboratory (2026) . The culmination of four years and four months of research, system architecture design, model innovation, infrastructure engineering, and production deployment at MARS. This laboratory symbolizes the transformation of MARS into a research-driven AI technology organization with internationally recognized contributions.
Figure 2.1: Overall architecture of MARS integrating quadtree-based representation with transformer refinement.
| Damage Category | Instances |
|---|---|
| Cracked Paint | 273,121 |
| Dent | 332,342 |
| Loose | 114,345 |
| Scrape | 434,237 |
Table 2.1: Thai Car Damage Dataset Statistics
| Method | Backbone | AP | AP50 | AP75 | APs | APm | APl | FPS |
|---|---|---|---|---|---|---|---|---|
| Mask R-CNN He et al. (2017) | R50-FPN | 31.7 | 50.1 | 34.7 | 11.9 | 29.9 | 41.3 | 8.4 |
| PointRend Kirillov et al. (2020) | R50-FPN | 33.9 | 51.7 | 36.4 | 12.3 | 31.0 | 42.2 | 4.6 |
| Mask Transfiner Ke et al. (2022) | R50-FPN | 34.9 | 52.4 | 37.1 | 13.8 | 32.5 | 45.0 | 6.7 |
| MARS (Ours) | R50-FPN | 36.2 | 53.0 | 38.9 | 15.8 | 34.6 | 47.3 | 6.8 |
Table 2.2: Instance Segmentation Performance Comparison
Figure 2.2: Comparison of segmentation results against SOTA methods.
Figure 2.3: Robust Multi-Scenario Damage Segmentation Performance of ALBERT. Qualitative results across diverse vehicle types, lighting conditions, occlusions, and damage complexities. ALBERT demonstrates strong boundary adherence, high-confidence instance classification, and effective discrimination between real structural damage and visually similar artifacts. Notably, the model maintains precise mask localization even under complex curvature surfaces and reflective materials, highlighting its readiness for production-grade automotive inspection systems.
Figure 2.4: Fine-grained mask boundary refinement achieved by MARS.
Figure 2.5: Fine-Grained Boundary Refinement and Small-Damage Sensitivity. Comparison across challenging small-scale damages including scratches, paint cracks, minor dents, and reflective distortions. ALBERT achieves superior mask precision with reduced background leakage and improved structural consistency compared to baseline approaches. The results confirm strong performance in small-object regimes, which are traditionally difficult yet critical in insurance claim validation and fraud detection workflows.
Figure 2.6: Qualitative comparison between ALBERT-v8 and ALBERT-v9. Version 9 shows improved boundary precision, better small-damage localization, and stronger structural consistency between predicted vehicle parts and damage types.
| Category | #Instances | Category | #Instances |
|---|---|---|---|
| scrape | 326,200 | missing | 23,790 |
| dent | 136,607 | sticker | 24,257 |
| loose | 94,995 | chip | 1,413 |
| crackedpaint | 74,799 | fake | 6,291 |
| torn | 31,725 | fakemud | 3,181 |
| scratch | 4,526 | fakeshadow | 11,954 |
Table 2.3: ALBERT-DAMAGE Dataset Statistics. Large-scale fine-grained vehicle damage segmentation dataset comprising 856,226 annotated instances across 26 damage categories. The dataset captures structural damage, surface-level defects, and hard-negative visual artifacts to enable robust real-world deployment.
| Category | #Instances | Category | #Instances |
|---|---|---|---|
| frontbumper | 22,432 | doorhandle | 26,054 |
| rearbumper | 18,726 | gastank | 5,663 |
| hood | 17,940 | frontpillar | 16,129 |
| frontfender | 25,998 | rearpillar | 12,508 |
| rearfender | 15,795 | rockerpanel | 15,310 |
| frontdoor | 24,086 | bedsidepanel | 8,396 |
Table 2.4: ALBERT-PART Dataset Statistics. Comprehensive structural vehicle part segmentation dataset containing 595,563 annotated instances across 61 fine-grained automotive components. The dataset covers exterior panels, lighting systems, glass regions, accessories, and structural elements, supporting large-scale production inspection systems.
| Model | AP | AP50 | AP75 | AP s | AP m | AP l |
|---|---|---|---|---|---|---|
| ALBERT (Damage) | 36.440 | 60.592 | 37.627 | 21.760 | 30.878 | 49.488 |
Table 2.5: Overall Instance Segmentation Performance of ALBERT (Damage Model)
| Category | AP | Category | AP | Category | AP |
|---|---|---|---|---|---|
| scrape | 20.372 | dent | 26.810 | loose | 14.585 |
| crackedpaint | 26.186 | torn | 13.451 | scratch | 27.995 |
| crack | 23.171 | brokenlight | 46.593 | crackedglass | 54.830 |
| shatteredglass | 61.567 | eartorn_1 | 30.952 | eartorn_2 | 31.500 |
| crush | 38.551 | missing | 44.191 | ding | 25.191 |
| ruined | 63.553 | sticker | 43.729 | chip | 54.098 |
Table 2.6: Per-Class Segmentation AP of ALBERT (Damage Categories)
| Model | AP | AP50 | AP75 | AP s | AP m | AP l |
|---|---|---|---|---|---|---|
| ALBERT (Part) | 62.317 | 85.737 | 68.460 | 33.102 | 56.122 | 73.815 |
Table 2.7: Overall Instance Segmentation Performance of ALBERT (Part Model)
| Category | AP | Category | AP | Category | AP |
|---|---|---|---|---|---|
| frontbumper | 69.302 | rearbumper | 79.510 | hood | 82.982 |
| frontfender | 66.774 | rearfender | 66.266 | frontdoor | 75.483 |
| reardoor | 77.840 | trunklid | 66.457 | frontwindshield | 78.282 |
| rearwindshield | 75.307 | frontsidewindow | 72.570 | rearsidewindow | 70.198 |
| sidewindow | 64.325 | sidemirror | 69.005 | headlight | 74.452 |
| grill | 59.826 | lowergrill | 52.100 | taillight | 72.027 |
Table 2.8: Per-Class Segmentation AP of ALBERT (Vehicle Part Categories)
Figure 2.7: Qualitative segmentation results of the ALBERT Part Model on diverse vehicles from the MARSAIL dataset. The model demonstrates strong capability in identifying multiple structural components including bumpers, doors, windshields, and lighting elements under real-world imaging conditions.
Figure 2.8: Additional examples highlighting the robustness of ALBERT for fine-grained vehicle component segmentation across diverse vehicle categories and viewpoints.
Figure 2.9: ALBERT accurately segments complex vehicle structures including grills, mirrors, and side panels while preserving sharp mask boundaries.
Figure 2.10: Examples illustrating stable segmentation performance across varying vehicle geometries including sedans, pickup trucks, and SUVs.
Figure 2.11: The ALBERT model successfully captures both large vehicle structures and smaller accessories such as door handles and logos.
Figure 2.12: Qualitative results demonstrating robust segmentation under varying illumination and background complexity.
Figure 2.13: Fine-grained segmentation results highlighting accurate delineation of adjacent vehicle components such as bumpers, grills, and headlights.
Figure 2.14: ALBERT maintains consistent part-level predictions across diverse viewpoints and occlusion patterns.
Figure 2.15: Examples showing reliable segmentation of overlapping structural components in complex real-world scenes.
Figure 2.16: Precise boundary localization of vehicle parts supports reliable downstream reasoning for damage localization and repair estimation.
Figure 2.17: Further qualitative examples illustrating ALBERT’s strong multi-scale feature representation for vehicle component understanding.
Figure 2.18: ALBERT consistently identifies vehicle components across varying camera distances and perspective distortions.
Figure 2.19: Robust segmentation across multiple vehicle body structures including roof components, pillars, and side panels.
Figure 2.20: Examples illustrating strong structural consistency in predicting complex component layouts across different vehicle designs.
Figure 2.21: ALBERT demonstrates stable segmentation performance even in challenging visual environments with cluttered backgrounds.
Figure 2.22: Qualitative damage segmentation results produced by the ALBERT Damage Model on the MARSAIL dataset. The model accurately detects diverse damage patterns including dents, scratches, cracks, and shattered glass across multiple vehicle surfaces.
Figure 2.23: Additional qualitative results demonstrating the robustness of ALBERT in detecting subtle surface damage across different vehicle colors, materials, and lighting conditions.
Figure 2.24: Examples illustrating the capability of ALBERT to localize fine-grained damage structures such as hairline cracks and small dents with high boundary precision.
Figure 2.25: ALBERT effectively identifies multiple co-occurring damage categories within a single vehicle image, supporting reliable multi-instance damage assessment.
Figure 2.26: Qualitative examples showing strong detection of structural deformation such as crushed panels and severely damaged surfaces.
Figure 2.27: The model successfully detects damage across a wide range of vehicle viewpoints, demonstrating strong generalization capability.
Figure 2.28: Examples illustrating reliable segmentation of glass-related damage such as cracked and shattered windshields.
Figure 2.29: ALBERT captures subtle surface defects including scratches and chipped paint, which are traditionally difficult to detect automatically.
Figure 2.30: Qualitative examples demonstrating consistent mask localization for complex and irregular damage patterns.
Figure 2.31: The model maintains strong performance even when damage appears under challenging environmental conditions such as reflections or shadows.
Figure 2.32: Additional examples showing ALBERT’s ability to capture both small cosmetic damage and large structural defects across multiple vehicle panels.
Figure 2.33: Robust qualitative results highlighting the scalability of ALBERT across diverse vehicle models and surface materials.
Figure 2.34: Examples illustrating the model’s capability to maintain high segmentation quality for overlapping and adjacent damage regions.
Figure 2.35: ALBERT accurately differentiates between genuine structural damage and visually misleading artifacts that could otherwise lead to incorrect insurance assessments.
Figure 2.36: The model consistently captures complex deformation patterns across multiple vehicle body panels.
| Method | IC15 | SVT | IIIT5K | SVTP | CUTE80 |
|---|---|---|---|---|---|
| RES50-ViT | 51.08 | 84.85 | 67.17 | 76.28 | 56.94 |
| RES50-DEF-(L3–L4) | 50.22 | 82.23 | 65.30 | 72.25 | 51.39 |
| RES50-ViT | 53.01 | 85.16 | 71.03 | 77.05 | 61.81 |
| RES50-ATT | 42.90 | 71.87 | 58.80 | 59.84 | 44.44 |
| RES50-ATT-ViT | 53.88 | 85.78 | 70.93 | 78.14 | 62.50 |
| RES50-ViT-PE | 57.05 | 87.33 | 73.33 | 81.09 | 65.97 |
Table 3.1: Performance comparison on IC15, SVT, IIIT5K, SVTP and CUTE80 datasets
Figure 3.1: Performance evaluation of the proposed DOTA-OCR model on the VIN recognition task (January–June 2025 test set). The model achieved an overall accuracy of 50.27% across 1,319 samples. The distribution of correct and incorrect predictions reflects the intrinsic difficulty of alphanumeric VIN recognition under real-world automotive and insurance imaging conditions, including metallic reflections, low contrast engraving, blur, and viewpoint distortion.
Figure 3.2: Performance evaluation of the proposed DOTA-OCR model on the Mileage recognition task (January–June 2025 test set). The model achieved an overall accuracy of 87.57% across 1,319 samples. A total of 1,155 predictions were correct, while 164 samples were incorrectly recognized. The error distribution highlights challenges inherent to odometer digit recognition in real-world automotive imagery, including glare from instrument clusters, motion blur, low illumination, partial occlusion, and varying dashboard designs.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
| Instance Type | g5.12xlarge |
|---|---|
| GPU | 4x NVIDIA A10G (24GB) |
| vCPU | 48 |
| Memory | 192 GB |
| Storage | EBS gp3/io2 (1-2 TB) |
| Typical Use | LoRA / QLoRA (7B-13B models) |
Table A.1: Instance Specification for Parameter-Efficient Fine-Tuning
| Instance Type | p4d.24xlarge |
|---|---|
| GPU | 8x NVIDIA A100 (40GB) |
| vCPU | 96 |
| Memory | 1152 GB |
| Networking | 400 Gbps (EFA) |
| Typical Use | FSDP / Multi-GPU Distributed Training |
Table A.2: Instance Specification for Distributed Fine-Tuning
| Instance Type | p5.48xlarge |
|---|---|
| GPU | 8x NVIDIA H100 (80GB) |
| vCPU | 192 |
| Memory | 2 TB |
| Networking | 3200 Gbps (EFA) |
| Typical Use | 70B+ or Multimodal Foundation Models |
Table A.3: Instance Specification for Foundation-Scale Training
| Objective | Action | Expected Outcome |
|---|---|---|
| Decouple Models | Convert each model to independent API service | Service-level scalability |
| Introduce Orchestrator | Implement lightweight agent controller | Dynamic stage execution |
| Logging Upgrade | Structured reasoning logs | Traceable AI decisions |
Table A.4: Phase 1 – Modular AI Refactoring
| Objective | Action | Expected Outcome |
|---|---|---|
| Vector Memory | Deploy embedding-based retrieval | Similar case reasoning |
| Feedback Loop | Integrate human QC corrections | Continuous improvement |
| Experience Replay | Store failed predictions | Error-aware refinement |
Table A.5: Phase 2 – Agent Memory Integration
| Objective | Action | Expected Outcome |
|---|---|---|
| Planner LLM | Deploy reasoning LLM for workflow selection | Non-linear task execution |
| Tool Selection Agent | Enable dynamic tool invocation | Flexible processing |
| Explainability Engine | Auto-generate reasoning trace | Regulatory compliance |
Table A.6: Phase 3 – Full Agentic Decision System