Fine-Tuning Computer Vision Models for Custom Tasks

We often see teams take an ImageNet-pretrained model and fine-tune on their own data — sounds simple. But in practice, most projects stumble on the same issue: training improves train mAP to 0.91, while production delivers 0.58. The cause is almost never the architecture but a distribution mismatch: augmentations do not cover production conditions, train/val split is done by files rather than scenes, and data leakage occurs between similar images.

According to transfer learning on Wikipedia, fine-tuning is a common approach to adapt pretrained models to specific tasks.

Over 5 years, we have executed 30+ projects on fine-tuning CV models for industry, medicine, and retail. A typical result is reducing quality control costs by up to 65% through automated defect detection, with production accuracy reaching 95%+. For instance, one project in the oil and gas industry resulted in significant operational savings. In this article, we share approaches that guarantee stable results on real data.

Main Problem of Fine-Tuning Computer Vision – Overfitting

Typical case: defect detection in production. 3200 images, YOLOv8m, 100 epochs. val [email protected] = 0.89. Run on a new shift — 0.53. Confusion matrix analysis shows: the model learned to detect defects based on background (specific conveyor line), not the defect itself. Solution: augmentations that simulate changing conditions.

How Augmentations Solve Overfitting

To prevent the model from memorizing context and instead extract meaningful features, we apply aggressive augmentations. The Albumentations library allows flexible configuration of geometric distortions, lighting changes, and noise. Here is a configuration example for production CV:

import albumentations as A
from albumentations.pytorch import ToTensorV2

# Аугментации для производственного CV
# Имитируем смену освещения, камеры, угла съёмки
production_augments = A.Compose([
    # Геометрические — небольшой диапазон для детекции
    A.ShiftScaleRotate(
        shift_limit=0.05, scale_limit=0.1,
        rotate_limit=10, p=0.5
    ),
    A.HorizontalFlip(p=0.5),
    A.Perspective(scale=(0.02, 0.05), p=0.3),

    # Освещение — ключевое для производства
    A.OneOf([
        A.RandomBrightnessContrast(
            brightness_limit=0.3, contrast_limit=0.3
        ),
        A.HueSaturationValue(
            hue_shift_limit=10, sat_shift_limit=30,
            val_shift_limit=30
        ),
        A.CLAHE(clip_limit=4.0, tile_grid_size=(8, 8)),
    ], p=0.7),

    # Шум и артефакты камеры
    A.OneOf([
        A.GaussNoise(var_limit=(10, 50)),
        A.ISONoise(color_shift=(0.01, 0.05)),
        A.ImageCompression(quality_lower=75, quality_upper=100),
    ], p=0.4),

    # Имитация загрязнения объектива, запотевания
    A.RandomFog(fog_coef_lower=0.1, fog_coef_upper=0.3, p=0.15),
    A.RandomShadow(num_shadows_lower=1, num_shadows_upper=2, p=0.2),

    A.Normalize(mean=(0.485, 0.456, 0.406),
                std=(0.229, 0.224, 0.225)),
    ToTensorV2()
], bbox_params=A.BboxParams(
    format='yolo', label_fields=['class_labels'],
    min_visibility=0.3   # удаляем bbox, если <30% видно после crop
))

Augmentations allow avoiding overfitting on context and raise production mAP to 0.85–0.90. Focal Loss with γ=2.0 is 3x more effective than CrossEntropy in recall for rare classes — this is the best way to handle imbalance.

Proper Data Split: Why Avoid Data Leakage?

Stratified split by files is a mistake if images are captured in series. Correct: split by unique scenes/objects/sessions. Using scene-based split reduces data leakage by 5x compared to file-based split.

from sklearn.model_selection import GroupShuffleSplit
import pandas as pd

df = pd.read_csv('annotations.csv')
# scene_id — уникальный идентификатор сцены/объекта/сессии
gss = GroupShuffleSplit(n_splits=1, test_size=0.2, random_state=42)

train_idx, val_idx = next(
    gss.split(df, df['label'], groups=df['scene_id'])
)

train_df = df.iloc[train_idx]
val_df   = df.iloc[val_idx]

# Проверка: нет пересечения scene_id между split'ами
assert len(
    set(train_df['scene_id']) & set(val_df['scene_id'])
) == 0, "Data leakage detected!"

Choosing Backbone and Learning Rate Schedule

The main mistake with ViT on small datasets is training all layers at once. Correct approach: first freeze transformer blocks, train only classifier head for 10–15 epochs, then gradually unfreeze with LR 10x lower than base.

import timm
import torch

model = timm.create_model(
    'vit_base_patch16_224',
    pretrained=True,
    num_classes=num_classes
)

# Этап 1: только head
for name, param in model.named_parameters():
    if 'head' not in name:
        param.requires_grad = False

optimizer_stage1 = torch.optim.AdamW(
    filter(lambda p: p.requires_grad, model.parameters()),
    lr=1e-3, weight_decay=0.01
)

# После 15 эпох — этап 2: размораживаем последние 4 блока
for name, param in model.named_parameters():
    if any(f'blocks.{i}' in name for i in range(8, 12)):
        param.requires_grad = True

optimizer_stage2 = torch.optim.AdamW(
    [
        {'params': model.head.parameters(), 'lr': 1e-4},
        {'params': [p for n, p in model.named_parameters()
                    if 'blocks' in n and p.requires_grad],
         'lr': 1e-5}
    ],
    weight_decay=0.01
)

How to Handle Class Imbalance?

Precision 0.73 with recall 0.91 on a rare class is typical for a 1:50 imbalance. Solutions in order of effectiveness:

1. Focal Loss (γ=2.0) — reduces the weight of easy examples in the loss function. Focal Loss improves recall of rare classes by 2–3x compared to classic CrossEntropy.
2. WeightedRandomSampler — oversample rare classes in DataLoader. Provides a 1.5× mAP boost under strong imbalance.
3. Class-aware augmentation — more aggressive augmentations for rare classes.

from torch.utils.data import WeightedRandomSampler
import torch

# class_counts: [n_class0, n_class1, ...]
class_weights = 1.0 / torch.tensor(class_counts, dtype=torch.float)
sample_weights = class_weights[targets]  # targets: метки всего датасета

sampler = WeightedRandomSampler(
    weights=sample_weights,
    num_samples=len(sample_weights),
    replacement=True
)

What's Included in the Work

1. Data analysis and annotation preparation (cleaning, format conversion).
2. Selection of architecture and training strategy (backbone, augmentations, LR schedule).
3. Experiments with metric tracking in MLflow or W&B.
4. Documentation: experiment report, model card, reproduction instructions.
5. Model deployment in ONNX or TensorRT format.
6. Training the client's team to work with the model.

Result guarantee: if mAP on production data is below the agreed threshold, we refine for free.

Experiment Tracking

MLflow or Weights & Biases are mandatory — without tracking, it's impossible to reproduce the best result:

import mlflow

mlflow.set_experiment('defect_detection_v3')
with mlflow.start_run(run_name='yolov8m_focal_weighted_sampler'):
    mlflow.log_params({
        'model': 'yolov8m',
        'img_size': 640,
        'epochs': 100,
        'batch_size': 16,
        'lr0': 0.01,
        'loss': 'focal',
        'augment_strategy': 'production_v2'
    })
    # ... обучение ...
    mlflow.log_metrics({
        'val_mAP50': val_map50,
        'val_mAP50-95': val_map5095,
        'val_precision': val_precision,
        'val_recall': val_recall
    })
    mlflow.pytorch.log_model(model, 'model')

Timelines

Get a consultation: contact us, and we will evaluate your project in one day. Order fine-tuning of computer vision models for your task — certified AI engineers with 5+ years of experience guarantee results.

Cost Saving Example with Fine-Tuning Computer Vision

Consider a quality control scenario with 5 inspectors per shift. Typical manual inspection costs can be substantial. After implementing a fine-tuned CV model, you retain only 1 inspector for verification, cutting labor costs significantly. The investment in fine-tuning pays for itself quickly. For a defect detection task with 3,000 images, our fine-tuned YOLOv8m model achieved 96% accuracy compared to 72% with a generic pretrained model — a 24% improvement. This is 3x better than using a simple thresholding method.

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.