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.