A road image can contain parked cars, moving vehicles, partially visible pedestrians, and objects reflected in glass. Before generating a batch of labels, decide which of those objects belong in your dataset. The useful starting point is a short annotation policy that two people can apply consistently.
LabelOp provides general image annotation tools for this work. The workflow below uses its existing labeling, AI, review, and export capabilities; it does not assume a driving simulator or an autonomous-driving evaluation system.
Define the objects before drawing them
Start with a small class list such as car, truck, bicycle, and pedestrian. Write down how to handle a rider on a bicycle, a person behind a windshield, or a vehicle cut off by the image edge. Decide whether a box should describe the visible portion of an object or an estimated full extent, and apply the same rule across the dataset.
Use boxes when your intended output is object detection. Choose polygons or click-assisted segmentation when your task requires an object's visible outline. Avoid adding detailed classes that annotators cannot reliably distinguish in the source images.
Try assistance on a varied sample
In LabelOp, local detection and zero-shot detection can provide annotation candidates. With a supported prompt-based model, enter the object names you want to find and inspect the results. Include a few difficult examples in the trial: small pedestrians, overlapping vehicles, dark scenes, and unusual viewpoints.
WebGPU accelerates click segmentation and Select Similar on compatible devices. Select Similar works from an object in the same image; it is not a promise of cross-frame tracking. Review candidates against your class policy before saving them.
Scale the draft, then review it
Managed Cloud AI offers detection and classification batches. Select an image range, set the confidence threshold, and decide whether to skip images with existing annotations. A completed batch means results were processed, not that every object is correctly labeled.
Assign images to teammates and use annotation review to approve work or request changes. Check both incorrect labels and missed objects. A high-confidence car prediction does not tell you whether a distant pedestrian was overlooked.
Check the handoff
Export a sample in the format your training pipeline expects. Open the output and check class names, image references, and geometry before exporting the full dataset. The Dataset Health Report can surface class distribution and geometry findings from supported annotation files; selected files are sent to LabelOp for processing.
Keep related road images together when planning evaluation splits so that near-identical views do not make the test misleading. That grouping is a dataset preparation decision, not automatic sequence-aware splitting in LabelOp.
Continue into training when ready
Export your annotations for your own tooling, or connect your own Vast.ai account to manage supported training jobs in LabelOp. Vast.ai GPU charges are separate from managed Cloud AI labeling. Start with a reviewed sample and verify the training configuration before expanding the run.