How many images do you need to train a custom YOLO dataset

You are teaching YOLO one thing it has never seen: your bracket, your tote, your cap. The question that decides your week is how many pictures that takes. This guide commits to numbers: an image budget per class, the resize rule that decides whether your captures are usable, a turntable capture plan that produces the budget in an afternoon, and the arithmetic that tells you when to raise the training resolution instead of shooting more. It assumes you fine-tune weights pretrained on COCO, which is what current YOLO trainers do by default, and it is written for the engineer who has to plan a capture day and defend the dataset size afterwards.

The image budget, by situation

The middle row is the default answer for one unique object: 200-300 images per class, and no fewer than about 100 even in the friendliest setup. The rule that keeps these budgets honest is to count images that add a condition, not files. A 30 fps clip of one turntable revolution is 900 frames but only about 30 distinct poses; count files and you will believe you have a dataset when you have a slideshow.

Resolution: the resize rule, not the source file

Every YOLO trainer letterboxes whatever you give it onto a square network size, 416x416 or 640x640 by default, and 640-1280 px on the long edge is the sensible starting range. Source resolutions therefore need not match: a 4000 × 3000 capture and a 1280 × 720 phone video train side by side. What the network actually sees is your object at its size after that resize, and that is the number to engineer.

The first line is the same arithmetic as calculating object-space pixel size, with the field of view standing in for magnification.

Rule of thumb: keep the object above roughly 30 px at the training resolution. The basis is grid geometry — the finest detection grid at 640 px works in 8 px cells, so an object smaller than about four cells has too few left to carry texture. Under that floor, narrowing the field of view or raising the training size fixes it; more images do not.

If a scene only works above imgsz 1280, the capture geometry is the problem, not the dataset; small pixel camera selection explains why tiny pixels do not rescue a wide field. Downscaling a mixed set yourself before training, as post-capture downscaling covers, is bookkeeping for other tools, not a YOLO requirement.

Turntable video: the fastest way to cover a full pose range

Frames cut from video are legitimate training images, and a turntable plus one video produces the 200-300 budget in an afternoon. The trap is correlation. Adjacent video frames are near-duplicates, and a dataset of near-duplicates trains a model that fails on the first pose it did not see. Space the extraction so consecutive frames differ, and vary the camera, not just the turntable.

The 1-5 s extraction spacing used in practice works for the same reason: at any ordinary turntable speed it puts at least 10° between poses, enough that successive frames stop predicting each other. Split by session, not randomly: a random split scatters neighboring frames on both sides of the divide and reports an accuracy your deployed model will not see.

Spend the budget on variation, not duplicates

A workable first split of 300 images for one object:

Read the first trained model as a data-collection tool: its false negatives name the images to shoot next. Include empty scenes in the set — images with no target teach the background and cut false positives, which is why background images carry no labels. Lighting is the variation engineers skip and later regret, because a line under skylights produces a different image at 9 a.m. and 4 p.m.; controlling ambient light variation covers fixing it on the line or baking it into the dataset.

Training settings that match the budget

Keep annotations in one format from the start; a CSV from one tool and YOLO text files from another is a conversion bug waiting to happen, and converting CSV annotations to YOLO walks the mapping. Then:

• Split by capture session: four sessions for training, the fifth for validation.
• Train 300 epochs as the first run, the standard starting point, and stop early once validation mAP sits flat for tens of epochs.
• imgsz 640 first; move to 1280 only when the resize arithmetic puts the object under the 30 px floor.
• When accuracy stalls below target, read the misses and capture images of those exact conditions. More epochs on the same images cannot teach a condition the dataset never showed.

If the run crashes before epoch one finishes, that is the environment, not the dataset; the usual GPU memory and CUDA segmentation faults have their own fixes. The wider training toolbox lives on the systems topic page.

Capture hardware that keeps the dataset small

The turntable plan needs steady frame delivery with no dropped frames, and enough capture resolution to keep the object over the 30 px floor after resize. The LS-CMR-0166 captures 4000 × 3000 at 30 fps over USB3 — the capture behind the worked example above.

Small objects that push you to imgsz 1280 reward sensor pixels, because pixels bought on the sensor survive the letterbox: the MV-CS200-10GC resolves 20 MP onto a 1" sensor with a C-mount.

To deploy the finished detector without the training PC, move inference to a smart camera: the MV-SC5060XM-08M-WBN pairs a 6 MP sensor with an 8 mm M12-mount lens over GigE.

Whichever camera captures the data, freeze exposure and gain in manual — auto exposure re-brightens each scene differently and silently adds a variable you never budgeted for, the failure behind image brightness drift. On a line, strobed lighting makes every session's images interchangeable; the light controller finder filters for strobe output.

Before you annotate: a five-minute pre-flight

• Zoom to 100% on ten sampled frames; edges crisp, no double edges from motion blur.
• Exposure and gain in manual, identical settings in every session.
• Resize arithmetic done for every scene type: object width after resize above 30 px.
• Extraction spacing honored: no two neighboring video frames both made the cut.
• At least 20-30 negative images with no target captured.
• The validation session copied to its own folder before any training, so it cannot leak in.
• Boxes tight to the object, one format, class names consistent.

The decision, per case

Fixed cell, one pose, controlled light: 100 images fine-tuned from pretrained weights, and stop there — the budget buys robustness, and a fixed scene has none to buy. One unique object seen from many angles and places: capture 200-300 with the turntable plan and let the first model nominate the next 15. Production across lines, seasons or many appearances: budget 1,000-2,000 per class and split training by environment. Object under 30 px after resize: no dataset size fixes it — narrow the field of view or train at 1280. A class added months later: retrain on the combined dataset, because fine-tuning only on new-class images erodes the classes the model already knew.

Size your own numbers with the resolution calculator above, pick the capture camera in area scan cameras or smart cameras, start the chain in the camera configurator, or bring the scene to us through the contact page.