Model reference — Pictures
Pose estimation locates a person’s joints in an image — shoulders, elbows, wrists, hips, knees, ankles — and returns them as points, giving a skeleton that describes how the body is positioned.
From those points, useful things follow: whether someone is standing, sitting or has fallen; whether a lift is being performed with a bent back; whether a worker has reached into a machine; how an athlete’s technique changes across a movement. It is also less intrusive than it appears, because a skeleton can be kept and the image discarded, which allows behaviour to be measured without retaining pictures of identifiable people. Accuracy depends on how much of the body is visible: occlusion by machinery, furniture or other people is the main source of error.
Models in this group take a single image as input: a photograph, a scan or a screenshot. 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 photograph at about 2 megapixels.
Vendor: Shanghai AI Laboratory (OpenMMLab)
What it does: locates joints quickly and accurately for several people in the same image, engineered for production throughput. The usual default.
| 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 (images/hour) | ≈ 25,000 | ≈ 90,000 | ≈ 260,000 |
Vendor: University of Sydney
What it does: the most accurate option, particularly where limbs are partly hidden or bodies overlap. Heavier per image, so it is used where precision matters more than volume.
| 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 (images/hour) | ≈ 6,000 | ≈ 24,000 | ≈ 65,000 |
Vendor: Carnegie Mellon University
What it does: the long-established model, able to track hands and feet in detail as well as the main skeleton. Slower, but well understood and widely referenced.
| 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 (images/hour) | ≈ 3,000 | ≈ 12,000 | ≈ 32,000 |
Vendor: Google
What it does: runs on device with no GPU, which is what allows pose to be measured on a camera or phone without sending any image anywhere.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 4 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 8 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (images/hour) | ≈ 40,000 | ≈ 150,000 | ≈ 450,000 |
Vendor: Ultralytics
What it does: detects people and their joints in one pass, which keeps a single model where you also need person counting or tracking from the same images.
| 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 (images/hour) | ≈ 30,000 | ≈ 120,000 | ≈ 340,000 |
The right model depends on how many people appear at once, how much of them is visible, and whether estimation runs on a server or on the camera. Our consultants review your images and what you need to measure, then recommend a model and the joint set to track.
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 Pose estimation (human). 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.