AI DEVELOPMENT / DEDICATED BUILD
Computer Vision
Infer useful structure from images or video through sampling, geometry, learned features, and task-specific evaluation.
What it is
Infer useful structure from images or video through sampling, geometry, learned features, and task-specific evaluation.
This page separates the concept, implementation path, interactive experiment, evaluation criteria, failure modes, and sources so the topic can be studied without relying on a summary tile.
Core concepts
- Images are sampled signals; resolution, noise, color space, and lens geometry affect what information is available.
- Convolution combines local neighborhoods with reusable kernels.
- Modern vision systems often learn features with convolutional networks or vision transformers.
- Classification, detection, segmentation, tracking, and pose estimation require different labels and metrics.
- Augmentation can improve robustness only when transformations preserve the intended label.
Architecture / workflow
- Define the visual task and acquisition conditions.
- Create train/validation/test splits that avoid near-duplicate leakage.
- Establish an image-processing or small-model baseline.
- Train and evaluate across lighting, viewpoint, occlusion, and device changes.
- Inspect failure examples visually rather than relying only on aggregate metrics.
Technical foundations
- Discrete convolution computes weighted local sums over pixel neighborhoods.
- Detection metrics incorporate localization quality as well as class prediction.
- Intersection over Union: IoU=area(pred∩truth)/area(pred∪truth).
- Image normalization and augmentation alter the effective training distribution.
Evaluation checklist
- Class confusion and per-class recall.
- IoU or average precision for localization tasks.
- Robustness by lighting, camera, scale, blur, and occlusion.
- Latency and memory on the actual deployment device.
- Manual review of false positives and false negatives.
Failure modes and limits
- Background correlations can dominate the object of interest.
- Random frame splits can leak nearly identical video frames.
- Model performance can collapse under lighting or camera changes.
- Visual confidence does not prove semantic understanding.
Practical build path
- Start with an explicit image-processing baseline.
- Train a small classifier with a clean split.
- Add augmentation and test each transformation separately.
- Extend to detection or segmentation only when the simpler task and evaluation pipeline are reliable.
Primary / authoritative references
AUTHORED STUDY COMPANION / 3.5
Inspect a spatial filter and its boundary assumptions
Prerequisites. Arrays, convolution/correlation and image sampling.
A local image filter combines nearby intensities with a spatial kernel. The notebook applies a horizontal Sobel-style correlation kernel to a synthetic 7×7 vertical step edge. Valid correlation produces a 5×5 output because it uses only positions where the entire 3×3 kernel fits.
The kernel sums to zero, so adding a constant brightness offset leaves the interior response unchanged. This algebraic property does not imply robustness to contrast scaling, clipping, noise, rotation or boundary padding. The operation is explicitly correlation: the kernel is not flipped as in the mathematical definition of convolution. Deep-learning libraries often use this convention while naming the operation convolution.
An edge response is not object recognition. A trained classifier additionally needs a dataset, target labels, split policy and task-level evaluation. Preserve this simple filter as a diagnostic baseline when studying learned features.
| Element | Model | Interpretation |
|---|---|---|
| Input | 7×7 step image | Synthetic fixture |
| Kernel | Zero-sum 3×3 weights | Local contrast response |
| Output | 5×5 valid map | Boundary pixels omitted |
Predict, run, compare
The vertical step gives maximum response 4. Adding 0.3 to every input pixel changes the valid response only by floating-point rounding.
Download executable notebooks and source Inspect source ↗
Open vision.ipynb in the package. Run cells in order, compare with the reference, then change the stated input. Restart the kernel to reset.
Computed reference result — local CPU fixture
{
"edge_map": [
[
0.0,
4.0,
4.0,
0.0,
0.0
],
[
0.0,
4.0,
4.0,
0.0,
0.0
],
[
0.0,
4.0,
4.0,
0.0,
0.0
],
[
0.0,
4.0,
4.0,
0.0,
0.0
],
[
0.0,
4.0,
4.0,
0.0,
0.0
]
],
"maximum_response": 4.0,
"brightness_shift_residual": 0.0,
"boundary": "valid correlation; no padding, image normalization, trained detector, or object recognition"
}This is the recorded result of the included code under the package environment. It is not a live service or device measurement.
Extend the experiment
Compare valid, zero-padded and reflected boundaries in copied code. Rotate the edge and describe which orientation the selected kernel responds to.
Check your reasoning: The unchanged response under an additive brightness shift follows from:
- Zero kernel sum — Correct. The constant term cancels when the weights sum to zero.
- A semantic understanding of the image — No semantic model is present.
- A large training set — The kernel is fixed and untrained.
This is an explained self-check, not a certification of mastery.
Claim-specific sources
Reviewed 2026-10-02: this worked example and its explicit assumptions. Source links and notebook results do not establish production or hardware validation.