Computer Vision: рынок

Вакансия

открытая вакансия, прошедшая гейт качества и содержащая не менее двух технологий

Наблюдаемый кандидат

кандидат, чей стек содержит эту технологию; агрегат, без единой записи наружу

Подходящий кандидат

пока не считается

Окно

открытые вакансии на момент сборки снимка

Топ-5: python, pytorch, c++, opencv, tensorflow. Python deep (монопольно — edge components иногда C++/Rust). PyTorch + torchvision mastery — 90%+ CV-research на PyTorch 2026 (TensorFlow legacy). OpenCV mastery: classical CV ops (cv2.findContours / cv2.HoughLines / Canny edge / morphology / homography / camera calibration / stereo matching — must для preprocessing + traditional algorithms). Pillow (PIL — image manipulation). Self-supervised + foundation models: DINOv2 (Meta — best vision backbone 2024 для feature extraction), CLIP + OpenCLIP + SigLIP (vision-language alignment), MAE (Masked Autoencoder). Classification backbones: ViT (Vision Transformer) family — ViT-B/L/H, Swin Transformer, ConvNeXt (modern CNN), EfficientNet + MobileNet (efficient), timm (PyTorch Image Models — Ross Wightman — pretrained models, must-library). Data augmentation: albumentations (industry standard — fast + comprehensive — 30+ augmentation types), torchvision.transforms.v2 (modern PyTorch native), RandAugment + AugMix + Mixup + CutMix (research recipes). Multimodal LLM: GPT-4V + Claude 3.5 Sonnet vision + Gemini 1.5/2.0 Pro vision + open-source LLaVA + Qwen 2 VL + InternVL + Idefics2 + Molmo (best open multimodal 2024). OCR: PaddleOCR (Baidu — multilingual leader 2026 — 80+ languages), EasyOCR, Tesseract (legacy still huge), Docling (IBM 2024 — doc understanding), Surya (multilingual modern), TrOCR (Microsoft transformer-based). Face recognition: InsightFace (open-source SOTA 2026 — ArcFace + RetinaFace + Buffalo models), DeepFace, face_recognition (dlib wrapper). Russian: NtechLab FaceNGN / VisionLabs Luna / Tevian / RecFaces. Video understanding + tracking: ByteTrack (fastest tracker 2026), DeepSORT, BoT-SORT, StrongSORT, MOTRv2. Video models: VideoMAE, X-CLIP, InternVideo, MViT. 3D vision: NeRF (original 2020) + Instant-NGP (NVIDIA — 100× faster) + Gaussian Splatting (3DGS) (rising 2023–2026 — фотореалистичный rendering, заменяет NeRF), Nerfstudio (unified framework), Open3D (point cloud processing), PyTorch3D (Meta), Kaolin (NVIDIA), COLMAP (structure-from-motion), Mitsuba 3 (differentiable rendering). Generative image / video: Stable Diffusion family (SDXL + SD3 + SD3.5 — Stability AI), FLUX.1 (Black Forest Labs — открытый рекорд качества 2024+), DALL-E 3 / Imagen 3 / Midjourney v6 / Ideogram (text-in-image). Video: Runway Gen-3 / Pika 1.5 / Kling AI / Sora / AnimateDiff / SVD. UIs для generative: ComfyUI (node-based — research-favorite), Automatic1111. Fine-tuning generative: LoRA + DreamBooth для Stable Diffusion / FLUX. Inference serving: Triton Inference Server + BentoML + TorchServe. Annotation: CVAT (OpenCV — industry standard для detection / segmentation), Label Studio + V7 Darwin (commercial) + Roboflow (rapid prototyping). Datasets: COCO (detection / segmentation benchmark) + ImageNet (classification) + OpenImages + LAION-5B (web-scale generative training) + Objects365 + Visual Genome + ADE20K (segmentation) + Cityscapes (autonomous driving). Hardware-aware optimization: quantization (INT8 / FP16), pruning, distillation, structured sparsity (for NVIDIA Ampere+ tensor cores).

Все три роли overlap в 2026, но focus areas различные. ML Engineer — generalist, владеет production ML stack (recsys / fraud / ranking / classical ML + LLM). Может работать с CV-data, но не deep specialization. См. ML Engineer. AI Engineer / LLM Engineer — focus на LLM integration (text-focused). См. AI / LLM Engineer. NLP Engineer — focus на text + speech. См. NLP Engineer. Computer Vision Engineer (эта страница) — focus на image / video / 3D data. Stack overlap с ML Engineer ~60% (PyTorch + cloud ML + deployment) + ~30% unique (OpenCV + Ultralytics + Detectron2 + SAM + albumentations + edge deployment specifics — TensorRT / CoreML). Distinctions: 1) Image processing intuition — CV Engineer understands camera models / lens distortion / color spaces / histogram analysis / morphological ops — classical CV foundation. AI Engineer обычно ничего такого не знает. 2) Computational constraints — CV models heavy (gigabytes), inference latency critical (autonomous driving — 30+ FPS mandate), edge deployment (TensorRT / CoreML / OpenVINO mastery) — exclusive CV territory. 3) Generative image / video specialization — Stable Diffusion / FLUX / ComfyUI workflow expertise — CV-niche specifically (overlaps с AI Engineer для text-driven generation). 4) 3D vision skills — NeRF / Gaussian Splatting / point cloud / SLAM / camera calibration — exclusive CV. 5) Domain expertise — CV roles часто tied к specific industry (autonomous vehicles / medical imaging / robotics / AR/VR / satellite imagery / manufacturing QC). Each domain имеет свои датасеты + правила. Career-pivots: ML Engineer Senior → CV Engineer — 4-8 месяцев (need OpenCV + Ultralytics + Detectron2 + albumentations + один domain). CV Engineer Senior → ML Engineer — 2-4 месяца (easy lateral, добавить MLOps + LLM basics). CV Engineer Senior → AI Engineer (LLM track) — 3-6 месяцев. Hot 2025–2026 sub-specializations: 3D Vision (Gaussian Splatting) / Generative Image-Video (FLUX + Sora-style) / Multimodal (Vision-Language Models — LLaVA / Molmo).

Decision tree для object detection / segmentation 2026: 1) Ultralytics YOLO (YOLOv8 / v9 / v10 / v11) — default choice 2026 для object detection. Pros: easy training (3-5 lines of code), good docs + community, fast inference (real-time на CPU + edge), wide deployment support (ONNX + TensorRT + CoreML + TFLite native). Cons: Ultralytics license requires AGPL or commercial license ($$) для proprietary use, less academic-flexible чем Detectron2 / MMDetection. Use case: 90% production object detection use cases, прототипы, edge deployment. 2) Detectron2 (Meta) — research-grade. Pros: clean API, flexible (easy custom architectures), Apache 2.0 license (commercial-friendly), backed by Meta. Cons: slower training than Ultralytics, larger learning curve. Use case: complex custom architectures, research projects, when you need fine control over training loop. 3) MMDetection / MMSegmentation / MMPose (OpenMMLab — Chinese consortium) — huge ecosystem. Pros: 100+ pre-implemented architectures (старые и новые), papers reference implementations always там first, Apache 2.0. Cons: config-heavy (steep learning curve), Chinese-language community sometimes hard для English-only, dependencies могут конфликтовать. Use case: research, comparing many architectures, paper reproduction. 4) DETR family (transformer-based detection) — modern paradigm shift 2020+. RT-DETR (Real-Time DETR — Baidu, rising 2024+ — competitor к YOLO в real-time space), DINO-DETR, Co-DETR. Pros: end-to-end (no NMS post-processing), often better accuracy on complex scenes. Cons: slower inference чем YOLO, more compute-hungry для training. Use case: highest-accuracy needs, complex scenes (crowded objects). 5) SAM (Segment Anything Model — Meta 2023) + SAM 2 (video version 2024) — universal segmentation. Use case: a) zero-shot segmentation (segment anything без training), b) interactive annotation tool (point / box prompt → mask), c) data labeling acceleration (SAM-assisted annotation 10× faster). Combine с YOLO / Detectron2: detector → SAM для precise masks. 6) Classical detection (HOG + cascade classifiers — OpenCV) — only для very simple cases, edge devices без GPU, low-power microcontrollers (still relevant для embedded scenarios). Default 2026 рекомендации: Production deployment + commercial → Ultralytics YOLO (если ok с AGPL or commercial license) OR Detectron2 (Apache 2.0 alternative). Research / paper reproduction → MMDetection. Highest accuracy / complex scenes → RT-DETR или Co-DETR. Universal / zero-shot segmentation → SAM 2. Annotation acceleration → SAM + interactive workflows (CVAT integration). Edge / mobile → YOLO с TensorRT (NVIDIA) / CoreML (Apple) / TFLite (Android). Senior CV Engineer должен знать когда какой инструмент.

Да, 25% CV Engineer-вакансий — full-remote или гибрид. CV work fully cloud-based (training в cloud GPUs — A100 / H100, datasets streaming с S3 / GCS, deployment в Kubernetes). Аутсорсеры — почти всегда remote на US-CV projects. Российские — гибрид или remote после probation. Российские банки (Сбер AI Banking CV — face recognition / document verification) — гибрид/офис security compliance. Autonomous vehicle companies — особый случай: research roles often remote-friendly, но product deployment / on-vehicle testing требует on-site (Pittsburgh Cruise / Mountain View Waymo / Palo Alto Tesla / etc). Generative AI companies (Stability / Black Forest Labs / Runway / Pika / Kling) — full-remote standard. International voice-AI / NLP companies overlap (multimodal teams) — full-remote. Big Tech CV — гибрид-standard. Релокант-хабы для CV: США (Bay Area + Pittsburgh — autonomous vehicle clusters + Boston — Robotics MIT region), UK (London — Wayve), Канада (Toronto — Vector Institute), Германия (Berlin AI + Munich automotive — BMW / Mercedes AI), Франция (Paris — Hugging Face + Mistral для vision LLM), Япония (Tokyo — Sony AI + automotive). Английский для international CV-remote — must (CVPR / ICCV / ECCV papers + community англоязычные).

3D Vision Engineer — sub-specialization внутри CV focused на 3D understanding + neural rendering. Day-to-day: 1) NeRF / Gaussian Splatting reconstruction — input video → 3D scene representation. Tools: Instant-NGP (NVIDIA — fast NeRF) / Nerfstudio (unified framework) / gsplat (Gaussian Splatting library) / Polycam app для capture. 2) Point cloud processing — LiDAR data (autonomous vehicles) или depth cameras (Kinect / RealSense). Tools: Open3D / PCL (Point Cloud Library) / PyTorch3D. 3) Structure-from-motion (SfM) — multiple 2D images → 3D scene. Tools: COLMAP (industry standard SfM). 4) SLAM (Simultaneous Localization and Mapping) — robotics / AR — track camera pose + build map. Tools: ORB-SLAM3 / OpenVSLAM / Kimera. 5) Differentiable rendering — learn 3D from 2D supervision. Tools: Mitsuba 3 / nvdiffrast / Kaolin. 6) 3D generation — text → 3D mesh (DreamFusion + Magic3D + Zero-1-to-3) или image → 3D (TripoSR / InstantMesh — open-source 2024+). 7) Mesh processing — texturing / retopology / UV unwrapping для 3D content pipelines. Stack-specific: PyTorch3D (Meta — 3D research), Kaolin (NVIDIA — 3D DL), Mitsuba 3 (differentiable rendering), Open3D / PCL (point cloud), COLMAP (SfM), Nerfstudio + gsplat (NeRF / 3DGS), Blender Python API (mesh manipulation). Career-flow: CV Engineer Senior + computer graphics interest + NeRF / 3DGS hands-on portfolio → 3D Vision Engineer — 6-12 месяцев.

Roadmap: 1) Math foundations — линейная алгебра + матанализ + основы projective geometry (transformations / homographies / camera models). «Multiple View Geometry in Computer Vision» Hartley / Zisserman — bible для CV math (можно использовать как reference, не читать целиком). 2) Python deep + ML basics — PyTorch + NumPy + Matplotlib. Build simple ML classifier (MNIST). 3) OpenCV mastery — classical CV ops. Курс: «Computer Vision Course» от PyImageSearch (Adrian Rosebrock — best entry для OpenCV), OpenCV official documentation + tutorials. Build pet-projects: edge detection / homography panorama stitching / face detection с classical Haar cascades. 4) Deep Learning для Vision — Stanford CS231n «Convolutional Neural Networks for Visual Recognition» (Karpathy / Li — free YouTube + slides — must-do, canonical CV deep learning course). 5) PyTorch + torchvision hands-on — train ResNet / EfficientNet on CIFAR-10, fine-tune pretrained model на own dataset. Учиться использовать timm library (pretrained models). 6) Object detection — Ultralytics YOLO hands-on (easiest entry — train YOLO on custom dataset за день). Затем Detectron2 (more flexible). 7) Segmentation — Mask R-CNN через Detectron2, затем SAM (universal segmentation) — try interactive segmentation. 8) Annotation tooling — CVAT mastery (industry standard для CV annotation). Annotate own dataset (10-50 images), train detector on it. 9) Augmentation mastery — albumentations library (must для production CV). Understand training stability — strong augmentation improves robustness. 10) Modern transformers для CV — Vision Transformer (ViT) + Swin Transformer + DINOv2 (foundation model). Hugging Face Transformers vision support. 11) Multimodal LLM — try GPT-4V + Claude vision + open-source LLaVA / Qwen 2 VL для understanding. 12) Generative CV track (popular 2024–2026): Stable Diffusion + FLUX hands-on, ComfyUI workflows mastery, LoRA / DreamBooth fine-tuning. Курс: «Generative AI with Diffusion Models» DeepLearning.AI. 13) 3D Vision track (rising 2024+): NeRF + Gaussian Splatting hands-on с Nerfstudio framework, capture own scene на phone (Polycam app), reconstruct as 3DGS. 14) Edge deployment hands-on — convert PyTorch model → ONNX → TensorRT (on NVIDIA GPU), benchmark inference latency. Try CoreML conversion для Apple Silicon. 15) Pet-проект portfolio: a) production-grade detection pipeline (e.g. fish-counting / vehicle-counting / cell-counting demo); b) custom Stable Diffusion / FLUX LoRA на own style; c) Gaussian Splatting reconstruction (cool 3D scene from phone video); d) mobile CV app (deploy via CoreML/TFLite). Document на GitHub + blog post + video demo. Курсы РФ: МФТИ DLSchool (CV модуль — free YouTube), Karpov.Courses «Computer Vision» track, Otus «Computer Vision», SkillFactory CV, School21 (Сбер) AI Computer Vision track. International (eng): Stanford CS231n (canonical — free YouTube), fast.ai Practical Deep Learning (CV included), Hugging Face Computer Vision Course (free), «Deep Learning for Computer Vision» book Mohamed Elgendy, PyImageSearch University (Adrian Rosebrock — applied focus). Books-must: «Deep Learning for Vision Systems» Mohamed Elgendy (Manning), «Multiple View Geometry» Hartley / Zisserman (math reference), «Computer Vision: Algorithms and Applications» Richard Szeliski (free 2nd edition online — encyclopedic). Conferences: CVPR (top — June), ICCV (October, alternating с ECCV), ECCV (October alternating), NeurIPS (December — broader ML с CV-track). Backend Senior / ML Engineer Middle + CV interest + portfolio → CV Engineer Junior — 6-12 месяцев. PhD CV / Robotics → Senior CV Engineer — direct entry.

222 активных открытых CV Engineer-вакансий с явной CV-спецификой в нашей выборке. Реальный пул многократно шире — many CV roles classified как general ML Engineer / Robotics / AI Engineer (titles типа «ML Engineer для autonomous driving» или «Senior Backend Engineer with CV focus»). True CV-focused dev jobs в СНГ + Европе оценочно 300–1500 позиций активных любой момент 2026 (с учётом нечётко классифицированных). География: 🇷🇺 Россия / 🇵🇱 Польша / remote. Реальный рынок шире за счёт международного remote-сегмента (generative AI companies + robotics startups full-remote-friendly). Время закрытия Senior CV Engineer — 6-12 недель (longer чем general AI Engineer за счёт rare-skill — PyTorch deep + classical CV + один domain + edge deployment combination).

Senior CV Engineer владеет полным циклом vision engineering + technical leadership. Math foundations: linear algebra + projective geometry (camera models / homographies / epipolar geometry) + calculus + optimization theory — на уровне «can read CVPR papers without math blocks». Python deep + Backend Senior level: async / typing / FastAPI / pytest mastery. C++ basics (для edge deployment + OpenCV custom kernels — nice to have). PyTorch + torchvision mastery deep: custom Datasets / Samplers / Losses / training loops, distributed training (DDP for multi-GPU CV training), mixed-precision (FP16 / BF16 mandatory для production), gradient accumulation для large models. OpenCV mastery: classical CV ops mastery (calibration / morphology / Hough transforms / homography / stereo matching / optical flow), C++ API basics для performance-critical paths. Modern transformers для CV: ViT / Swin / DINOv2 mastery, fine-tuning strategies, hybrid CNN-transformer architectures. Object detection / segmentation mastery: Ultralytics YOLO production deployment + Detectron2 custom architecture authoring + MMDetection / MMSegmentation when needed + RT-DETR + SAM 2 integration в pipelines. Data augmentation mastery: albumentations advanced (custom transforms + composition strategies), test-time augmentation (TTA), AutoAugment / RandAugment / CutMix / Mixup understanding. Foundation models: DINOv2 + CLIP + SigLIP + OpenCLIP — use cases для feature extraction + zero-shot classification + retrieval. Multimodal LLM (GPT-4V / Claude vision / LLaVA / Qwen 2 VL / Molmo) — when to use vs train custom model. Generative CV mastery (if track включает): Stable Diffusion + FLUX deep — ControlNet / LoRA / DreamBooth / IP-Adapter / textual inversion fine-tuning. ComfyUI workflow authoring. Video generation basics (AnimateDiff / SVD). 3D Vision mastery (if track включает): NeRF + Gaussian Splatting + Nerfstudio + COLMAP SfM + Open3D point clouds. Camera calibration + projective geometry deep. Edge deployment mastery: ONNX (cross-framework export + ONNX Runtime optimization), TensorRT (NVIDIA — INT8 quantization + plugin development если нужно), CoreML (Apple Silicon — Conv2D ops support + Neural Engine specifics), TFLite (Android NNAPI delegate), MediaPipe pipelines, OpenVINO (Intel iGPU). Hardware-aware optimization: quantization (PTQ + QAT), pruning, distillation (teacher-student), structured sparsity (NVIDIA Ampere+ для 2× speedup). Production CV pipeline architecture: design end-to-end pipeline на whiteboard — data ingestion (cameras / drives / cloud streaming) → preprocessing → batched inference → post-processing → tracking / aggregation → downstream actions. Multi-camera systems: synchronization + calibration + fusion для autonomous / surveillance. Inference serving: Triton Inference Server (NVIDIA — best для multi-model + dynamic batching), DeepStream (NVIDIA video analytics), BentoML / TorchServe. Latency budgets: design for real-time constraints (autonomous: 30+ FPS, AR: 60+ FPS, video moderation: batch ok), profile-driven optimization (NVIDIA Nsight + PyTorch Profiler). Domain expertise: deep understanding одного-двух domains (autonomous driving / medical imaging / retail visual search / robotics perception / AR/VR / satellite imagery / manufacturing QC) — главный premium-driver Senior+. Annotation strategy: design large-scale annotation workflows (CVAT / Label Studio + SAM-assisted), active learning loops, label quality control. Evaluation methodology: COCO metrics (AP / AP50 / AP75 / mAP@[.5:.95]) deep understanding, segmentation metrics (mIoU / Dice / boundary F1), domain-specific metrics (NDS / mAP для autonomous detection), human-in-the-loop evaluation для generative. Soft: ADRs writing для CV architecture decisions, technical writing (CV feature design docs + paper drafts если research-track), cross-team collaboration (Product / Backend / Robotics / Hardware teams), mentoring Middle CV Engineers, paper-reading discipline (CVPR / ICCV / ECCV / NeurIPS must-follow). Английский для Senior+ MUST — CV community / docs / papers / conferences англоязычные. Optional bonus: open-source contributions в Ultralytics / Detectron2 / MMDetection / timm / albumentations / Nerfstudio — резко повышают market value для Big Tech CV / autonomous vehicles / generative AI hiring. Papers at CVPR / ICCV / ECCV workshops — премиум для Research Engineer CV track.