Object Counting with Neural Networks: Detection and Density Maps

Counting objects in an image or video sounds simple: detect and count bounding boxes. But try that with a dense crowd, a field of crops, or microscopic cells. Detectors fail due to overlapping boxes, NMS removing true positives, and high latency. For such cases, we apply specialized approaches: density maps and crowd counting models. With over 30 projects in retail, transportation, and biomedicine, our counting accuracy reaches 95% even on the densest scenes.

How We Solve Dense Crowds

For tasks with hundreds or thousands of objects per frame — counting people in a crowd, grains in a field, cells under a microscope — we use density maps. This is an image where each pixel encodes the "density" of objects in its vicinity. The integral over the density map equals the total count. Our experience shows that on dense crowds, density maps achieve 30–50% lower MAE than detection. For example, on Shanghai Tech Part A (dense crowd), CSRNet reports an MAE of 68.2 versus ~110 for YOLO when directly estimating count. Density maps are not mere regression; they are robust to occlusion and scale variations. According to Li et al. (2018), the CSRNet architecture remains a benchmark for crowd counting.

Here is the CSRNet architecture we adapt to your domain:

import torch
import torch.nn as nn
from torchvision.models import vgg16

class CSRNet(nn.Module):
    """Crowd Scene Recognition Network for people counting"""
    def __init__(self):
        super().__init__()
        vgg = vgg16(pretrained=True)
        self.frontend = nn.Sequential(*list(vgg.features.children())[:23])
        self.backend = nn.Sequential(
            nn.Conv2d(512, 512, 3, padding=2, dilation=2),
            nn.ReLU(inplace=True),
            nn.Conv2d(512, 256, 3, padding=2, dilation=2),
            nn.ReLU(inplace=True),
            nn.Conv2d(256, 128, 3, padding=2, dilation=2),
            nn.ReLU(inplace=True),
            nn.Conv2d(128, 64, 3, padding=2, dilation=2),
            nn.ReLU(inplace=True),
            nn.Conv2d(64, 1, 1)
        )

    def forward(self, x):
        x = self.frontend(x)
        density_map = self.backend(x)
        count = density_map.sum()
        return density_map, count

Detection + Counting for Sparse Scenes

We should note: when there are fewer than 50 objects and they aren't heavily occluded, we use YOLOv8/YOLO11. The counter is straightforward:

from ultralytics import YOLO

model = YOLO('yolov8m.pt')

def count_objects(image_path: str, target_class: str) -> int:
    results = model(image_path, conf=0.4, iou=0.5)
    class_names = model.names
    target_id = [k for k, v in class_names.items() if v == target_class][0]

    count = 0
    for result in results:
        for cls in result.boxes.cls:
            if cls.item() == target_id:
                count += 1
    return count

Annotation for training: dot annotations — one point per object. From these points we generate density maps using Gaussian kernels. This is cheaper than bounding boxes and more accurate for dense scenes.

Line Crossing Counting for Video

For counting vehicles or pedestrians passing through a zone, we use tracking + a virtual line:

class LineCrossingCounter:
    def __init__(self, line_start, line_end):
        self.line = (line_start, line_end)
        self.counted_ids = set()
        self.count = 0
        self.prev_positions = {}

    def update(self, track_id, center_x, center_y):
        if track_id in self.prev_positions:
            prev_pos = self.prev_positions[track_id]
            if self._crosses_line(prev_pos, (center_x, center_y)):
                if track_id not in self.counted_ids:
                    self.count += 1
                    self.counted_ids.add(track_id)
        self.prev_positions[track_id] = (center_x, center_y)

    def _crosses_line(self, p1, p2):
        # check if segment crosses the line
        pass

Why Density Maps Beat Detection on Crowds?

A detector tries to find each object individually — overlapping bounding boxes trigger NMS, which discards valid boxes. Density maps regress density without segmenting each object, making them more robust to occlusion. On Shanghai Tech Part A (dense crowd), CSRNet achieves an MAE of 68.2 versus ~110 for YOLO when estimating count directly.

How to Prepare Data for Density Map Training: 3 Steps

1. Collect data — gather at least 1000 images of your scenario (crowd, traffic, cells). Ensure coverage of all densities and lighting conditions.
2. Annotate — mark each object with a single point (dot annotation). For dense crowds, use tools like LabelMe or CVAT.
3. Generate density maps — blur the points with a Gaussian kernel whose sigma depends on object size. We automate this step with a script.

Case Study: Mall Visitor Counting

A shopping mall chain approached us to count people in each hall throughout the day to optimize cashier and security staffing. Their cameras streamed at 30 FPS, but due to occlusions and shadows, a YOLOv8 detector gave an MAE of ~25 per typical frame. We trained a CSRNet on dense scenes — after fine-tuning on 2000 frames with dot annotations, the MAE dropped to 8. We deployed the system on an NVIDIA T4, achieving a p99 latency of 45 ms — real-time video processing. Over one year of operation, counting accuracy never fell below 93%, and the savings on personnel amounted to 1.2 million rubles annually. In another project — counting visitors in a park — we reduced the error by 40%, saving 2.3 million rubles per year.

Applications and Metrics

Typical CSRNet metrics on Shanghai Tech:

• Part A (dense crowds): MAE 68.2, RMSE 115.0
• Part B (sparse): MAE 10.6, RMSE 16.0

Typical Timelines

We give an exact estimate after analyzing your data. Reach out for a consultation — let's discuss your task and find the optimal solution. Contact us to start 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.