Custom AI for Sign Language Recognition: MediaPipe, Transformer, CTC

Development of AI for Sign Language Recognition

Deaf employees in manufacturing often lack real-time information – alert systems are not adapted to sign language. Attempts to deploy off-the-shelf translators fail due to high gesture variability and lack of support for Russian Sign Language (RSL). We build custom AI solutions that solve this problem at the engineering pipeline level. Our company specializes in computer vision and NLP; we developed a pipeline using MediaPipe for keypoint extraction and a Transformer for classification, achieving accuracy up to 84% on isolated signs and WER as low as 17% on continuous signing. We have implemented over 10 sign language recognition projects, accumulating more than 5 years of experience.

Why is Continuous Sign Language Recognition Harder than Isolated Signs?

An isolated sign is a single gesture from a vocabulary (e.g., HELLO). Continuous Sign Language Recognition (CSLR) is a stream of signs with undefined boundaries, where adjacent signs influence each other. CSLR requires CTC decoding (Connectionist Temporal Classification), which predicts a sequence of characters without segmentation. An additional challenge is motion-based disambiguation: signs identical in shape differ only in movement trajectory. Humans handle this with context; neural networks use BiLSTM layers to analyze temporal dependencies.

How We Use MediaPipe and Transformer for Isolated Signs

Our pipeline extracts 258 features per frame: 21 keypoints per hand + 33 body landmarks (MediaPipe Holistic). Normalization relative to shoulder distance provides scale invariance – a gesture from 0.5 m is recognized the same as from 2 m. The Transformer processes a window of 30 frames (1 second) and outputs class probabilities. The code below shows architecture and inference.

import numpy as np
import cv2
import torch
import torch.nn as nn
import mediapipe as mp
from dataclasses import dataclass
from collections import deque
from typing import Optional
import json

@dataclass
class SignPrediction:
    sign_id: int
    gloss: str          # sign name (HELLO, WATER, THANK_YOU)
    confidence: float
    hand: str           # left / right / both
    frame_range: tuple  # (start_frame, end_frame)

class HandLandmarkExtractor:
    """
    Extracts 21 keypoints per hand + 33 body landmarks from MediaPipe.
    Normalization relative to shoulder distance for scale invariance.
    """
    def __init__(self):
        self.mp_holistic = mp.solutions.holistic
        self.holistic = self.mp_holistic.Holistic(
            static_image_mode=False,
            model_complexity=1,
            min_detection_confidence=0.5,
            min_tracking_confidence=0.5
        )

    def extract(self, frame: np.ndarray) -> Optional[np.ndarray]:
        """
        Returns feature vector of dimension 258:
        - 21 right hand points × 3 (x,y,z) = 63
        - 21 left hand points × 3 = 63
        - 33 body points × 4 (x,y,z,visibility) = 132
        """
        rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
        results = self.holistic.process(rgb)

        # Right hand
        rh = np.zeros(63, dtype=np.float32)
        if results.right_hand_landmarks:
            for i, lm in enumerate(results.right_hand_landmarks.landmark):
                rh[i*3:i*3+3] = [lm.x, lm.y, lm.z]

        # Left hand
        lh = np.zeros(63, dtype=np.float32)
        if results.left_hand_landmarks:
            for i, lm in enumerate(results.left_hand_landmarks.landmark):
                lh[i*3:i*3+3] = [lm.x, lm.y, lm.z]

        # Body pose
        pose = np.zeros(132, dtype=np.float32)
        if results.pose_landmarks:
            for i, lm in enumerate(results.pose_landmarks.landmark):
                pose[i*4:i*4+4] = [lm.x, lm.y, lm.z, lm.visibility]

        features = np.concatenate([rh, lh, pose])

        # Normalization: shoulder distance as scale
        # Points 11 (left shoulder) and 12 (right shoulder) in pose
        left_shoulder = pose[11*4:11*4+2]
        right_shoulder = pose[12*4:12*4+2]
        shoulder_dist = np.linalg.norm(left_shoulder - right_shoulder)
        if shoulder_dist > 0.01:
            features[:126] /= shoulder_dist  # normalize only hands

        return features if (results.right_hand_landmarks or
                            results.left_hand_landmarks) else None


class SignLanguageTransformer(nn.Module):
    """
    Transformer for keypoint sequences.
    Input: (batch, seq_len, 258) – window of 30–60 frames.
    Trained on WLASL (2000 ASL signs) or PHOENIX-2014T (German).
    """
    def __init__(self, input_dim: int = 258,
                 d_model: int = 256,
                 nhead: int = 8,
                 num_layers: int = 4,
                 num_classes: int = 2000,
                 dropout: float = 0.1):
        super().__init__()
        self.input_proj = nn.Linear(input_dim, d_model)
        self.pos_emb = nn.Embedding(300, d_model)

        encoder_layer = nn.TransformerEncoderLayer(
            d_model=d_model, nhead=nhead,
            dim_feedforward=d_model * 4,
            dropout=dropout, batch_first=True
        )
        self.encoder = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
        self.classifier = nn.Sequential(
            nn.LayerNorm(d_model),
            nn.Linear(d_model, d_model // 2),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(d_model // 2, num_classes)
        )

    def forward(self, x: torch.Tensor,
                 src_key_padding_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        """x: (B, T, 258)"""
        B, T, _ = x.shape
        positions = torch.arange(T, device=x.device).unsqueeze(0).expand(B, -1)
        x = self.input_proj(x) + self.pos_emb(positions)
        x = self.encoder(x, src_key_padding_mask=src_key_padding_mask)
        x = x.mean(dim=1)
        return self.classifier(x)


class SignLanguageRecognizer:
    """
    Real-time isolated sign recognition.
    Sliding window buffer + threshold-based sign start/end detection.
    """
    WINDOW_SIZE = 30   # 30 frames @ 30 fps = 1 second
    MIN_SIGN_FRAMES = 10

    def __init__(self, model_path: str,
                  vocabulary_path: str,
                  confidence_threshold: float = 0.7,
                  device: str = 'cuda'):
        self.device = device
        self.extractor = HandLandmarkExtractor()

        with open(vocabulary_path) as f:
            self.vocabulary = json.load(f)  # {id: gloss}

        self.model = SignLanguageTransformer(
            num_classes=len(self.vocabulary)
        ).to(device)
        checkpoint = torch.load(model_path, map_location=device)
        self.model.load_state_dict(checkpoint['model_state_dict'])
        self.model.eval()

        self.threshold = confidence_threshold
        self.frame_buffer: deque = deque(maxlen=self.WINDOW_SIZE)
        self.sign_active = False
        self.sign_start_frame = 0
        self.frame_count = 0

    def process_frame(self, frame: np.ndarray) -> Optional[SignPrediction]:
        self.frame_count += 1
        features = self.extractor.extract(frame)

        if features is None:
            if self.sign_active and len(self.frame_buffer) >= self.MIN_SIGN_FRAMES:
                return self._classify_buffer()
            self.frame_buffer.clear()
            self.sign_active = False
            return None

        self.frame_buffer.append(features)
        self.sign_active = True

        if len(self.frame_buffer) < self.WINDOW_SIZE:
            return None

        return self._classify_buffer()

    @torch.no_grad()
    def _classify_buffer(self) -> Optional[SignPrediction]:
        seq = np.stack(list(self.frame_buffer))  # (T, 258)
        tensor = torch.from_numpy(seq).unsqueeze(0).float().to(self.device)

        logits = self.model(tensor)
        probs = torch.softmax(logits, dim=-1).squeeze()
        conf, pred_id = probs.max(dim=0)
        conf = float(conf.item())
        pred_id = int(pred_id.item())

        if conf < self.threshold:
            return None

        gloss = self.vocabulary.get(str(pred_id), f'SIGN_{pred_id}')
        return SignPrediction(
            sign_id=pred_id,
            gloss=gloss,
            confidence=round(conf, 3),
            hand='both',
            frame_range=(self.frame_count - len(self.frame_buffer),
                         self.frame_count)
        )

How Our CSLR Pipeline Works

For continuous signing we use CNN + BiLSTM + CTC. Input – windows of 90 frames (3 seconds). CTC decoder outputs a sequence of glosses, removing repeats and blank symbols. Our experience shows that the key difficulty is fingerspelling (manual alphabet). Letters are shown sequentially, and standard CSLR often confuses them. We solve this with a separate lightweight classifier based on MediaPipe, which processes each frame and corrects the main decoder's output. This reduces WER by 5–7% in tests on PHOENIX-2014T.

Our Transformer approach is 1.5–2× faster than I3D with comparable accuracy. For CSLR we use an ensemble of BiLSTM with attention, yielding WER 3–5% lower than standard non-segmented CTC. According to WLASL-2000, Transformer achieves 84% on isolated signs.

What the Work Includes

• Requirements analysis: vocabulary definition (number of signs, specifics – RSL, ASL, BSL), data collection and annotation.
• Model training: architecture selection (Transformer / BiLSTM + CTC), augmentation, fine-tuning on the target dataset.
• Inference optimization: quantization (INT8), export to ONNX, p99 latency reduction to 30–50 ms.
• Integration: REST API, RTSP stream, desktop application.
• Documentation and training: handover of model card, instructions for adapting to new signs.

Project Workflow

1. Data audit – assess quality and sufficiency of video material (1–2 weeks).
2. Prototyping – baseline model on 50–100 signs (2–4 weeks).
3. Deployment – CI/CD, MLflow tracking, drift monitoring (2–3 weeks).
4. Iteration – fine-tuning on real customer data (4–8 weeks).

Indicative Timelines

Cost is calculated individually – depends on data volume, number of classes, and required accuracy. To evaluate your project, contact us – we will provide a preliminary plan and timeline within 2–3 days. We have been working with sign languages for over 5 years; our engineers are certified in MediaPipe and PyTorch. Get a consultation to discuss the details.

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.