Domain-Agnostic Object Counting, Detection and Tracking in Images
Overview
This service provides ready-to-use AI pipelines for detecting, counting, segmenting, and tracking objects in images or video across different application domains. It is intended for users who need flexible visual analysis without commissioning a fully custom model for each new task. The service relies on pre-trained models and focuses on configuring inference workflows to the user's data and operational scenario rather than training new models. Depending on the use case, the user supplies reference examples, prompts, or initialization inputs, and the service returns processed outputs such as counts, detections, masks, and tracked object trajectories. This makes it suitable for rapid deployment in cases where robustness and adaptability are needed across varying object types and environments. The service relies on pre-trained, domain-agnostic computer-vision models for few-shot object counting and detection and segmentation-based object tracking. The main methodologies used are GECO2 for few-shot counting and detection, and DAM4SAM for tracking by segmentation, with the service focused on configuring inference pipelines rather than training customer-specific models.
Outcome
Deliverables include configured inference pipelines based on pre-trained models and structured outputs for downstream use. Depending on the task, outputs include object counts, bounding boxes, confidence scores, segmentation masks, and tracking results over image sequences or video.
Service Card
| Category | AI products |
| Domain | Visual Analytics, Computer Vision |
| Vertical | Cross-domain |
| Target users | Industry, Research Organizations, Public Sector Users, Cross-Domain Analytics Teams |
| Service type | Model and Pipeline |
| Outputs | Configured inference pipelines using pre-trained models, with outputs such as object counts, bounding boxes, confidence scores, segmentation masks, and tracking results. |
| Requirements | Requires NVIDIA GPU node. Few-shot counting and detection use a small set of example annotations. Tracking requires an initial object region or mask together with image or video sequences for inference. |
| Limitations | |
| Availability | Q4/2026 |
| Licensing | MIT and CC BY-NC-SA 4.0 |
| Visual content | |
| Visual caption |