AI-Powered Car Damage Assessment from Photos

We developed an AI-powered visual car damage assessment system that solves real-time insurance scoring. Our engineers have over 10 years of experience in computer vision and have delivered more than 50 projects. The company has been in the AI solutions market for over 5 years. Imagine: a customer photographs a car in the parking lot, uploads 6 shots into an app, and within 40 seconds receives a preliminary damage estimate—no expert visit, no inspection queue. This architecture, based on YOLOv8-seg + regression cost model, can be deployed to production in 8–12 weeks. It guarantees a 30–50% reduction in expert costs. For a typical insurance portfolio, the budget savings can reach 2–3 million rubles per year; for a portfolio of 10,000 policies, savings amount to about 15 million rubles annually.

What the Model Actually Detects

The task breaks down into three levels:

• Localization of damage — where and what: dent, scratch, crack, broken glass, body deformation. A segmentation model outputs a polygonal mask for each defect with pixel area.
• Part classification — which car zone the damage belongs to: front bumper, hood, left front wing, door, etc. This is critical for mapping to the hourly rates of specific parts.
• Severity estimation — light / medium / severe damage. A regression head on backbone features provides a score that is multiplied by coefficients from the cost reference table.

Pipeline Architecture

A two‑stage scheme works more reliably than end‑to‑end:

1. Damage detection and segmentation → YOLOv8-seg or SAM 2 with automatic prompts per car zone
2. Part classification → ResNet-50 or EfficientNet-B3 trained on body breakdown (30–50 part classes)
3. Cost estimation → XGBoost or LightGBM on features: mask area, part class, severity score, make/model (from EXIF or a separate classifier)

# Example inference of the segmentation model
from ultralytics import YOLO
import cv2

model = YOLO('damage_seg_v8x.pt')  # fine-tuned on damage dataset
results = model.predict(
    source='car_damage.jpg',
    conf=0.35,
    iou=0.45,
    imgsz=1280,   # high resolution critical for small scratches
    retina_masks=True
)

for r in results:
    masks = r.masks.data       # (N, H, W) tensor
    classes = r.boxes.cls      # damage classes
    areas = masks.sum(dim=[1,2])  # area of each mask in pixels

How We Ensure Detection Quality

Why are small scratches the most common problem?

At standard imgsz=640, a 10 cm scratch at 1.5 m distance occupies 3–4 pixels. YOLOv8 misses them with IoU < 0.3. Raising imgsz to 1280 increases recall for small defects from 0.44 to 0.71, but latency grows from 28 ms to 95 ms on RTX 3090. We balance this by automatically selecting resolution based on camera metadata.

How does the model distinguish new damage from old?

Fresh chips have white edges, old ones have brown rust edges. We add color analysis of mask edges and cross-check with the history of previous inspections. This reduces legal risks for the insurance company.

Dataset and Training

Minimum dataset size for a working model: 5,000 annotated images with masks covering all damage and part classes. In practice we use:

• Open datasets: CarDD (4,000 images), COCO with fine-tuning
• Synthetic from Blender — for rare cases (total front destruction)
• Client data from insurance case archives

Fine‑tuning YOLOv8x-seg on RTX 4090 with 8,000 images takes about 18 hours (100 epochs, batch=16, imgsz=1280). Final metrics: [email protected] = 0.79, [email protected]:0.95 = 0.61.

Performance Comparison on Different Devices

Detection accuracy varies by damage type. For example, recall on dents reaches 0.85, while on small scratches it is 0.71.

What’s Included in the Work

• Data audit — checking dataset representativeness, recommendations on collection and labeling
• Model development — fine-tuning, quantization (INT8/INT4), export to ONNX/TensorRT
• Integration API — FastAPI endpoints for photo upload, mask retrieval, and cost calculation
• Mobile SDK — CoreML (iOS), TFLite (Android) with integration example
• Documentation — OpenAPI spec, model card, labeling instructions for new data
• Support — 3 months warranty after launch

How We Build the Damage Assessment Model: Step‑by‑Step Process

1. Collection and labeling — 5,000+ images with polygonal masks from insurer experts
2. Training — transfer learning from YOLOv8x-seg, 100 epochs, 8–12 hours on RTX 4090
3. Validation — testing on hold‑out set ([email protected] > 0.75) and on real cases from the client
4. Optimization — TensorRT FP16 for server, CoreML/TFLite for mobile
5. Deployment — Docker container on Kubernetes, monitoring p99 latency and drift

Common Model Errors and Their Solutions

• Lighting and glare — simulate RandomSunFlare in augmentations, quality check rejects overexposed photos
• Small scratches — increase imgsz to 1280 and use retina_masks, raising recall from 0.44 to 0.71
• Old vs new damage — color analysis of mask edges + cross-check with inspection history

Project Timelines

8 to 16 weeks depending on availability of labeled data and integration requirements with the insurer’s systems. Cost is calculated individually.

Contact us for a consultation on your project. Order a pilot project to evaluate accuracy on your own data. Get demo access to the system with your photos.

How Distribution Shift Kills CV Model Metrics in Industry

On a production line, a camera is installed to control product quality. The model is trained on 10,000 labeled images—test accuracy mAP 0.84. Deployed to production, and in the first week it misses 30% of defects. Lighting on the line changes between shifts; distribution shift nullifies the metrics. This is a classic story with computer vision in industry, where pattern recognition fails without proper drift handling.

Our engineers, with experience from 60+ computer vision projects, know how to eliminate such scenarios. We guarantee stable model performance under real conditions.

Object Detection: YOLO, RT-DETR, and Everything in Between

YOLO is the standard for real-time detection. YOLOv8 and YOLOv11 from Ultralytics are the most used versions in production: simple API, active community, built-in validation, and export to ONNX/TensorRT. For tasks with high accuracy requirements and less critical latency, RT-DETR, a transformer-based architecture without NMS, gives better mAP on COCO at comparable speed to YOLOv8l.

A typical mistake when training a detector: dataset of 8000 images, 3 classes, fine-tune YOLOv8m—F1 0.73 on validation. Look at confusion matrix—one class is almost never detected. Cause: imbalance 1:23. Solution: oversampling rare class, focal loss for objectness, augmentations (Mosaic, MixUp disabled for rare class as they "blur" it). Transfer learning is mandatory: pretrained on COCO weights reduces data requirement by 10 times. Fine-tuning on 500–2000 domain images yields a working model in 1–2 days on a single GPU.

For edge deployment: export to ONNX → TensorRT engine. YOLOv8n in TensorRT FP16 on Jetson AGX Orin gives 150+ FPS at P99 latency < 8 ms—3 times faster than ONNX Runtime without TensorRT. On server A10G: 700+ FPS for YOLOv8n in TensorRT INT8.

How Does Fine-Tuning YOLO Help in Pattern Recognition?

Suppose you need to find micro-defects on a metal surface—a task with high resolution and class imbalance. We use YOLOv8m pretrained on COCO and fine-tune on 2000 proprietary images. Apply augmentations Mosaic, MixUp, random perspective. After 200 epochs, mAP 0.5 reaches 0.93. Key techniques:

• Focal loss for the objectness head—reduces contribution of easily classified examples.
• Class-balanced sampling—equalizes representation of rare classes.
• Test Time Augmentation (TTA)—increases recall by 5–7% through averaging over flips and scales.

Get a consultation on architecture selection for your task—contact us.

Segmentation: SAM, Mask R-CNN, and Instance Segmentation

SAM (Segment Anything Model) from Meta changed the approach to segmentation. SAM 2 works with video, supports object tracking across frames—for interactive object selection by point or bbox, it's the best out-of-the-box choice. For production instance segmentation without interactive prompting, Mask R-CNN or YOLOv8-seg are used. YOLOv8-seg trains like a regular detector with additional masks, convenient in the same pipelines. Semantic segmentation (each pixel is a class) uses SegFormer, DeepLabV3+. SegFormer-B5 provides a good balance of accuracy and speed for satellite imagery or medical segmentation.

Case study: cell segmentation on microscopic images. Dataset of 400 images with manual annotation. Training Mask R-CNN on ResNet-50 backbone gave IoU 0.61—poor. Problem: objects (cells) overlap; standard NMS kills overlapping predictions. Solution: switch to cellpose (specialized architecture for biomedical tasks) + soft-NMS. IoU increased to 0.79.

OCR: When Tesseract Fails

Tesseract is a starting point for simple tasks: printed text, good lighting, straight layout. As soon as there are handwritten elements, non-standard fonts, perspective distortions, or multi-column layouts, Tesseract degrades quickly.

PaddleOCR is a production-grade solution: text block detection + recognition + structural analysis. Works out of the box for 80+ languages, including Russian. Supports tables and complex document structures. TrOCR (Microsoft) is a transformer OCR with strong results on handwritten text. For Russian handwritten text, fine-tuning is needed: the base model is trained mostly on Latin script.

What to Do When Tesseract Cannot Handle Pattern Recognition on Documents?

For tasks like "extract data from invoices/contracts/passports," we use LayoutLMv3 or Donut—these models understand document layout, not just text. Integration via Hugging Face Transformers, fine-tuning on 200–500 annotated documents. Typical pipeline:

1. Preprocessing: deskew, denoising, binarization via OpenCV.
2. Text block detection: PaddleOCR detection or CRAFT.
3. Recognition: PaddleOCR recognition or TrOCR.
4. Post-processing: normalization, validation via regex or LLM for structured fields.

For documents with fixed structure, template matching + OCR by coordinates is often more reliable than an end-to-end solution.

Face Recognition: Identification and Verification

Face recognition = detection + alignment + embedding + matching. Each stage matters.

Detection: RetinaFace or InsightFace for accurate face localization and keypoints. MTCNN is older but reliable. Embedding: ArcFace (InsightFace) is state-of-the-art for face recognition embeddings. Models iresnet50/iresnet100 pretrained on MS1MV3 (5M identities). Embedding vector 512 float32, comparison by cosine similarity. Threshold tuning: decision threshold is a critical parameter. At threshold 0.6, typical FPR on LFW benchmark is 0.001, TPR is 0.985. In production, threshold must be calibrated to the real distribution: people in masks, with changed appearance, different lighting conditions. Liveness detection is mandatory: MiniFASNet—lightweight model on CPU; FaceX-Zoo contains several pretrained liveness detectors.

Video Analytics

Video is a sequence of frames plus a temporal dimension. A naive approach—detecting on every frame—is expensive.

Tracking: ByteTrack and BoT-SORT are the standard for multi-object tracking. They work on top of any detector, adding persistent IDs to objects across frames—enabling object counting, motion tracking, velocity.

Optimization: not every frame needs processing. For static scenes, detect every 5–10 frames, with tracking in between. For event detection (person entering a zone), background subtraction (OpenCV MOG2) serves as a lightweight pre-filter before neural detection. Action recognition: SlowFast, VideoMAE for action classification. Heavy models—for production use ONNX export + TensorRT or offline processing.

How to Measure Pattern Recognition Model Quality in Production?

Quality monitoring is key to MLOps. We track:

• Prediction confidence distribution.
• Share of low-confidence predictions (indicator of OOD data).
• Drift of input images via feature distribution (embeddings from backbone).

A drop in average confidence from 0.87 to 0.71 over a week is an early signal of distribution shift. NVIDIA Triton Inference Server recommends tracking these metrics via Prometheus. Our certified engineers set up monitoring and guarantee SLA for inference quality.

Deployment of CV Models

For online inference, we use Triton Inference Server (NVIDIA)—production standard for serving CV models. Supports TensorRT, ONNX, PyTorch, dynamic batching, multiple instances. REST and gRPC API. We guarantee stable operation under load.

Edge deployment: ONNX Runtime on ARM/x86 CPU. TensorFlow Lite for mobile devices. OpenVINO for Intel CPU/GPU/VPU—gives 2–3× speedup on Intel hardware compared to ONNX Runtime. After deployment, we hand over the model with documentation and train personnel.

What Is Included in the Work

Deadlines and Cost

A prototype detector on existing data takes 1–2 weeks. Production system with optimization for target hardware takes 4–8 weeks. Full cycle including data labeling (1000–5000 images) takes 2–4 months. Cost is calculated individually for each task. Typical savings from implementing a quality control system can be significant per production line.

We have been in the market for over 5 years and completed 60+ computer vision projects. We will evaluate your project end-to-end—request a consultation to get a quote and technical proposal.