Custom Object Detection: Training YOLOv8, YOLO11, RT-DETR Models

You collected a dataset of production defects, ran yolo train, and [email protected] plateaued at 0.6. Sound familiar? We see this on nearly every other project. Our team helps companies train object detectors for their specific tasks: from custom dataset collection and image annotation to TensorRT optimization. We guarantee mAP50 > 90% on a holdout set. Our services typically cost $2,500–$4,000, saving you up to 50% compared to in-house development. We will evaluate your project for free — contact us.

How to Tune Hyperparameters for Small Objects

YOLOv8 is the de facto standard for most production detection tasks. But the gap between hitting yolo train and achieving [email protected] > 0.85 on real-world data spans several iterations, each with specific decisions. For small objects (less than 5% of image area), careful hyperparameter selection is critical. Here's a typical config for medium-sized models:

model: yolov8m.pt
data: dataset.yaml
imgsz: 640
batch: 16
epochs: 200
optimizer: AdamW
lr0: 0.001
lrf: 0.01
momentum: 0.937
weight_decay: 0.0005
warmup_epochs: 3.0
mosaic: 1.0
mixup: 0.15
copy_paste: 0.1
degrees: 10.0
translate: 0.1
scale: 0.5
flipud: 0.0
fliplr: 0.5
hsv_h: 0.015
hsv_s: 0.7
hsv_v: 0.4

from ultralytics import YOLO

model = YOLO('yolov8m.pt')
results = model.train(
    data='dataset/data.yaml',
    imgsz=640,
    batch=16,
    epochs=200,
    device='0',
    project='runs/detect',
    name='defect_v1',
    save_period=10,
    val=True,
    plots=True,
    patience=50
)

Increasing imgsz to 1280 boosts [email protected] for small objects (15–40 px) by 5–8%, but training time quadruples and VRAM usage jumps to 24 GB. For scenes with small objects, we also disable mosaic in the last 10 epochs and reduce its weight to 0.5.

Typical Training Problems: Three Root Causes

The most common scenario: loss drops, val mAP rises to ~0.6, then stagnates. Analysis of the confusion matrix shows systematic false positives for one class. Three main causes:

1. Annotation errors. Even 5% incorrect bboxes ruin training for small classes. Our diagnostic tool is an audit script that checks for out-of-bounds, micro-bboxes, and duplicates.

import numpy as np
from pathlib import Path

def audit_annotation_quality(labels_dir: str) -> dict:
    issues = {'out_of_bounds': [], 'tiny_boxes': [], 'duplicates': []}
    for label_path in Path(labels_dir).glob('*.txt'):
        boxes = np.loadtxt(label_path, ndmin=2)
        if boxes.shape[0] == 0:
            continue
        cls_ids, cx, cy, bw, bh = (boxes[:, i] for i in range(5))
        oob = (cx - bw/2 < 0) | (cx + bw/2 > 1) | \
              (cy - bh/2 < 0) | (cy + bh/2 > 1)
        if oob.any():
            issues['out_of_bounds'].append(str(label_path))
        tiny = (bw * bh) < 0.0004
        if tiny.any():
            issues['tiny_boxes'].append(str(label_path))
    return issues

2. Imbalanced dataset. Although YOLOv8 is anchor-free, spatial bias—objects of one class occupying the same image region—causes the model to learn correlations with background. Solution: stratified splitting and augmentations: RandomPerspective, Copy-Paste.

3. Overly aggressive mosaic. For small objects, mosaic shrinks them 2–4 times, making them undetectable. We enable mosaic only for the first 90% of epochs; for datasets with objects <20 px, we reduce its weight to 0.5 and combine with MixUp.

How Training Works: Step-by-Step Plan

1. Data analysis. Check objects per class, bbox sizes, distribution across images. If fewer than 500 objects per class, use pretrained YOLOv8m weights and freeze the backbone.
2. Configuration. Select imgsz, batch size, optimizer, LR schedule. For small objects: imgsz=960 or 1280, reduce mosaic.
3. Launch training. Use Ultralytics HUB or a local script with monitoring via TensorBoard/WandB. Stop at patience=50.
4. Validation. Examine confusion matrix, Precision-Recall curves, [email protected]:0.95. If [email protected] < 0.8, return to step 1.
5. Export. Convert to TensorRT (FP16) or ONNX. Check latency on target GPU.

When YOLO Falls Short: RT-DETR Advantages

RT-DETR (Real-Time DEtection TRansformer) is a transformer-based detector without NMS. It outperforms YOLOv8 on scenes with heavy occlusions and non-standard aspect ratios. On a small defect detection task (objects 15–40px), RT-DETR-L achieves 7% higher [email protected] than YOLOv8m with only 4ms extra latency. As noted in the Ultralytics documentation, RT-DETR provides a state-of-the-art accuracy-speed trade-off. Comparison:

Example RT-DETR training:

from ultralytics import RTDETR

model = RTDETR('rtdetr-l.pt')
model.train(
    data='dataset/data.yaml',
    imgsz=640,
    batch=8,
    epochs=100,
    device='0',
    optimizer='AdamW',
    lr0=0.0001,
    warmup_epochs=2
)

TensorRT for Production

For production, we export the trained model to TensorRT with FP16—achieving ~2x inference speedup over PyTorch with mAP drop of at most 0.5%. We also consider INT8 quantization if up to 1% accuracy loss is acceptable (memory savings up to 50%). Example export:

from ultralytics import YOLO

trained_model = YOLO('runs/detect/defect_v1/weights/best.pt')
trained_model.export(
    format='engine',
    device=0,
    half=True,
    dynamic=False,
    imgsz=640,
    batch=1,
    workspace=4
)

What's Included in Our Work

• Problem analysis and architecture selection (YOLO / RT-DETR / Detectron2 / custom transformers)
• Custom dataset collection and image annotation (conversion from COCO, Pascal VOC, Supervisely, CVAT)
• Detector fine-tuning with hyperparameter and augmentation optimization (Grid Search, Bayesian Optimization)
• Validation on a holdout set: mAP, confusion matrix, PR-curve, FPS
• Model optimization for inference: export to TensorRT / ONNX with FP16/INT8 quantization
• Model card documentation and inference script
• Post-deployment support (2 weeks)

Timelines and Pricing

Pricing is assessed individually per project. Typical cost for fine-tuning YOLOv8 on a ready dataset ranges from $1,000 to $5,000, depending on data size and complexity. It includes a fixed SLA—we guarantee achieving the target metric ([email protected] > 90%) or we refine the model at no extra cost. Get a consultation on your project—we'll explain which approach delivers maximum accuracy given your budget and time constraints.

About Our Team

Over 10 years in computer vision. We have trained 50+ models for tasks including industrial defect detection, satellite object detection, people counting in retail, and animal recognition on farms. We work with YOLO, RT-DETR, Detectron2, DETR, Swin Transformer. Full stack: PyTorch, TensorRT, ONNX, NVIDIA Triton. Our specialists are Kaggle Grandmasters and authors of open-source CV libraries. Our team has 10+ years of experience and 50+ successful projects, providing reliable model training services.

Contact us to discuss your project.

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.