Model reference — Video
Detection and tracking in video finds the objects in each frame and follows each one across frames, so that a vehicle, pallet, tool or animal has a single identity and a path through the scene rather than appearing anew in every frame.
Tracking is what makes video analysis different from applying an image model repeatedly. Once objects have paths, questions become answerable that a single frame cannot address: how many distinct vehicles passed, how long a pallet stood in a walkway, whether a person entered a restricted zone, how fast something moved. The tracker is usually the limiting component, and its main failure is identity switching — two objects crossing and swapping identities — which corrupts every count derived from it.
Models in this group take a recorded video file as input and are applied frame by frame. Each specification table gives three hardware tiers — Minimum, the smallest setup on which the model runs correctly; Medium, the usual production configuration; and High, a configuration sized for peak volume. The rate is the number of sample inputs processed per hour on that hardware. Use these rates for initial sizing. Before production, benchmark your own data to validate accuracy, latency, throughput and cost. The sample input here is one minute of 1080p video at 25 frames per second.
Vendor: Ultralytics
What it does: the standard detector for video: fast, accurate and available in several sizes. Fitted to your own object types, it is the starting point for nearly every tracking deployment.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (video minutes processed per hour) | ≈ 1,600 | ≈ 6,000 | ≈ 16,000 |
Vendor: Ultralytics
What it does: the previous generation, still widely used and supported by more third-party video tooling, which eases integration with existing camera systems.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (video minutes processed per hour) | ≈ 1,800 | ≈ 6,500 | ≈ 18,000 |
Vendor: Baidu (PaddlePaddle)
What it does: needs no duplicate-box removal step, giving cleaner detections and more predictable timing, which suits a real-time pipeline.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (video minutes processed per hour) | ≈ 700 | ≈ 2,600 | ≈ 7,000 |
Vendor: University of Science and Technology of China
What it does: better on small and overlapping objects, which is where tracking on wide industrial views usually fails.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (video minutes processed per hour) | ≈ 600 | ≈ 2,200 | ≈ 6,000 |
Vendor: Huazhong University of Science and Technology
What it does: follows objects between frames using their motion, which is fast, accurate and the usual default.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (video minutes processed per hour) | ≈ 1,200 | ≈ 4,000 | ≈ 10,000 |
Vendor: Tel Aviv University
What it does: adds appearance matching so an object is recognised after being hidden, which is what prevents identity switches in busy scenes.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (video minutes processed per hour) | ≈ 800 | ≈ 2,800 | ≈ 7,000 |
Vendor: community-maintained open-source project
What it does: the established tracker, lighter and well documented, adequate where objects are well separated.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (video minutes processed per hour) | ≈ 1,400 | ≈ 5,000 | ≈ 13,000 |
Vendor: Meta
What it does: outlines tracked objects precisely across frames, which is needed where size, area or shape must be measured rather than just position.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3090 | RTX 4090 | A100 80 GB |
| VRAM | 12 GB | 24 GB | 80 GB |
| vCPUs | 8 | 12 | 24 |
| RAM | 32 GB | 48 GB | 96 GB |
| Server | 1× RTX 3090 24 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (video minutes processed per hour) | ≈ 120 | ≈ 450 | ≈ 1,200 |
The right pairing depends on object density, how much objects occlude each other, and whether processing is on recorded files or live. Our consultants review your footage and what you need counted, then recommend the detector and tracker and report the identity-switch rate on your own material.
At the start of a project we may run a short proof of concept on a sample of your own data, measuring the accuracy and the throughput the model actually achieves on your material. That replaces the estimates on this page with real figures, so the cost and the schedule for the full engagement are known before it is committed.
Share a representative sample, expected volume, latency target and deployment location for General object detection and tracking. We benchmark the shortlisted models, recommend the lowest-cost GPU configuration that meets the target, and scale it from pilot capacity to a dedicated production cluster.