Model reference — Pictures
Object detection finds the things in an image and returns a labelled box around each one — a vehicle, a pallet, a person, a helmet, a specific product on a shelf — with a confidence score.
It is the workhorse of applied computer vision, and in practice almost every deployment is a fitted one: the general models know the few hundred everyday categories they were trained on, but the objects a business cares about are usually its own. Fitting a detector to your objects needs images of them with boxes drawn by hand — typically a few hundred to a few thousand per category — and that labelling effort, not the model, is what determines the timeline of a project. Once fitted, detection is cheap and fast enough to run on every image or every video frame.
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: Ultralytics
What it does: the current standard detector: fast, accurate and available in several sizes so the same model family covers a small camera device and a server. The default starting point for a fitted detector.
| 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) | ≈ 40,000 | ≈ 160,000 | ≈ 450,000 |
Vendor: Ultralytics
What it does: the previous version of the same family, still widely deployed and supported by more third-party tooling, which matters when integrating with existing equipment.
| 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) | ≈ 45,000 | ≈ 180,000 | ≈ 500,000 |
Vendor: Baidu (PaddlePaddle)
What it does: a detector that needs no post-processing step to remove duplicate boxes, which makes its output cleaner and its timing more predictable — useful in real-time systems.
| 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) | ≈ 18,000 | ≈ 70,000 | ≈ 200,000 |
Vendor: University of Science and Technology of China
What it does: a recent detector with better accuracy on small and overlapping objects, which is where most industrial detection struggles.
| 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) | ≈ 15,000 | ≈ 60,000 | ≈ 170,000 |
Vendor: Microsoft
What it does: detects and describes objects without being fitted to your categories first, which makes it the quickest way to establish what is achievable before any labelling is commissioned.
| 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) | ≈ 1,800 | ≈ 6,000 | ≈ 14,000 |
Vendor: Meta
What it does: older two-stage detectors that are slower but sometimes more accurate on large, well-separated objects, and are still the reference implementation in several regulated industries.
| 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) | ≈ 4,000 | ≈ 16,000 | ≈ 45,000 |
Model choice depends on object size, how many appear per image, the accuracy you need and whether it runs on a server or a camera. Our consultants review your images and categories, estimate the labelling required, and recommend a model with the accuracy it will reach on your objects.
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. 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.