Small Objects and Fine Details: Resolution Requirements

What this specification is for

Small parts, fine print and thin features fail first when image resolution is wrong. A screw head, a connector pin, a printed lot code or the tip of a pipette disappears if the camera is too far or the lens too soft, and no amount of labeling brings it back. Because the cost of re-collecting is high, the capture specification should be written from the recognition requirement backwards.

Existing archives rarely help with this, since their cameras were chosen for another job; the purpose-built route is covered in existing assets versus new collection. For related object families see fasteners and small hardware and electronic components recognition. The cluster overview is the product and object data hub.

Method: start from the feature, not the camera

1. Name the decision. Is the model deciding class, instance, orientation, defect or the legibility of a mark? Each needs a different feature size.
2. Pick the smallest feature that decides it. A hex socket for a screw type, the gap between pins, a character height in a printed code.
3. Set pixels across that feature. This is your choice, and depends on the detector, the accuracy needed and whether an upstream crop is used.
4. Fix the working distance range. Use the distance your robot's cameras will actually have, including approach and hold poses, not the distance that is convenient for the capture crew.
5. Back out optics. Ground sample distance (the size in the scene covered by one pixel) is roughly working distance times pixel pitch divided by focal length. Check this against the pixels-across-feature target.
6. Test motion and focus. Depth of field narrows as magnification grows, and a moving hand or arm adds blur.

Illustrative worked arithmetic (inputs are arbitrary, not minimums)

Every number below is invented to show the calculation. Do not read them as recommended values for any task.

# Illustrative only. Inputs are arbitrary example values, not minimums.
smallest_deciding_feature_mm = 2.0       # e.g. a hypothetical socket width
pixels_across_feature_chosen = 8         # buyer's choice for this example
required_px_per_mm           = 8 / 2.0   # = 4.0 px/mm
required_gsd_mm_per_px       = 1 / 4.0   # = 0.25 mm/px

working_distance_mm          = 400       # example
pixel_pitch_mm               = 0.003     # example sensor
focal_length_mm              = 4.0       # example lens
achieved_gsd_mm_per_px       = 400 * 0.003 / 4.0   # = 0.30 mm/px

# 0.30 > 0.25 -> this example setup misses its own target.
# Options: move closer, longer focal length, smaller pixel pitch, or relax the target.

The point is the check, not the figures. The target and the achieved footprint should both appear in the capture plan, and each delivered image batch should be tested against them.

Illustrative capture parameters

Another illustrative block; values are synthetic placeholders a buyer would replace.

Sensors and the published reference

T-LESS illustrates how capture design for small industrial parts can be documented. It covers 30 industry-relevant texture-less objects, many of them symmetric or mutually similar, with three synchronized sensors (structured-light RGB-D, time-of-flight RGB-D and high-resolution RGB), approximately 39K training and 10K test images per sensor, and a CAD and reconstructed 3D model for each object (Hodan et al., 2017).

It is a capture-design reference for pose estimation of texture-less parts. It does not report which resolution a small-part recognizer needs, and it says nothing about your parts. Use it to see how multi-sensor capture, CAD models and train/test separation were handled, not to infer a number.

Choose sensors for the failure you expect: high-resolution RGB for fine print and edges, depth sensors for pose and clearance, polarized or diffuse lighting for shiny parts. Depth sensors resolve less lateral detail than a high-resolution color camera at the same distance, so check them separately.

Views and diversity

Resolution is wasted if views are narrow. Include several orientations per part, resting states (flat, on edge, in a bin), pose in hand, partial occlusion by fingers or fixtures, and mixed bins where similar parts touch. For viewpoint planning see the multi-view product photography spec, and for pose labels, 6D pose annotation.

Acceptance tests

• Pixels-on-feature audit. Measure a sample of delivered images against the target and report the distribution, not an average.
• Blur check. Reject frames where the deciding feature is not sharp, using a rule you define.
• Metadata check. Distance, lens and exposure present on every frame.
• Label audit. A sample reviewed by someone who understands the parts, with disagreement reported.
• Split rule. Hold out part lots or sites, not random frames from one session.
• Consent and site access. Any people, workplace markings or confidential parts handled as the agreement states. This is not legal advice.

What higher resolution cannot do

More pixels do not remove ambiguity between parts that differ only by material, weight or tolerance. They do not show how a screw seats, how a connector clicks or how a small part slips between fingers. Handling small objects needs robot or teleoperation data, with hand or gripper observations, actions and contact information; recognition images do not stand in for it. Higher resolution also raises storage, bandwidth and annotation cost, and may exceed what an onboard model can process in real time, so decide whether to crop or downsample before training.

Where parts are transparent or reflective, resolution is not the main problem; see transparent and reflective objects.

Lab consumables are a typical small-and-transparent case; see laboratory consumables and instruments for the object-family view.

Questions to ask suppliers or capture partners

1. Which cameras, lenses and sensor pitches will be used, and can you supply a calibration report?
2. How will pixels across the deciding feature be measured and reported per batch?
3. What working distances and angles can the setup reproduce?
4. How is motion blur controlled for handheld or in-hand capture?
5. Can raw, unprocessed images be delivered alongside any compressed versions?
6. Who labels the parts, and what is their experience with these items?
7. How are confidential or proprietary parts handled in the images?
8. What rights do the part owner, photographer and site grant?
9. Can a separate lot or site be captured for evaluation?

Describe your requirement

If a small-part perception project needs purpose-built imagery, describe the parts, feature sizes, viewing distances and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. If no existing asset fits, SourceX can scope a commissioned collection of new recordings. A request does not guarantee a match; SourceX follows up about potential sources. The physical-world collection page describes the separate path for new recordings, and the humanoid training data guide shows the other routes.

Discuss your humanoid training-data requirements with SourceX

Sources

• T-LESS: An RGB-D Dataset for 6D Pose Estimation of Texture-less Objects — multi-sensor capture of texture-less industrial parts with CAD models; a capture-design reference.