Computer vision learning path

Computer vision studies what can be inferred from visual measurements and which assumptions make that inference possible. This route starts with sampled images and projective geometry, not with a list of model names. It then separates recognition, dense prediction, representation learning, generation, temporal and 3D perception, and evaluation so that a new paper has a durable scientific home.

Latest literature discovery audit: 2 October 2026. Existing paper reviews retain their individual source-check dates. A year in a paper table reports publication status; it is not a quality ranking or an invitation to make dated “update” sections.

Image restoration and inverse problems were included in the targeted discovery audit on 2 October 2026; no new paper was promoted in this pass.

Every detailed chapter ends with a catalog-generated Complete maintained paper route; its shorter staged table is the teaching spine, not the full bibliography.

What you should know first

Complete the optimization, neural-network, and evaluation portions of the ML and deep-learning route. You should be able to:

• manipulate tensors with batch, channel, height, width, token, and time axes;
• derive the output size of convolution, pooling, and patchification;
• distinguish training, validation, internal test, and external test data;
• read cross-entropy, regression, contrastive, and reconstruction objectives;
• measure an experiment under a fixed data, compute, and adaptation budget.

The open Szeliski textbook, the MIT Vision Book, and Stanford CS231n are useful references. Course notes explain mechanics; the staged tables in this site lead to primary papers.

The map

Stage	Maintained chapter	Central question	Exit evidence
1	Image foundations and recognition	How do sampling, geometry, convolution, and inductive bias turn an image into a prediction?	derive a receptive field and compare classical, CNN, and transfer baselines fairly
2	Detection and segmentation	How do models localize objects, pixels, instances, and text-specified concepts?	diagnose classification, localization, boundary, and prompt failures separately
3	Representation and foundation models	Which supervision, invariance, target, and scale create transferable visual features?	compare frozen, shallow, and full adaptation without changing the evidence budget
4	Generative vision	How do likelihood, latent variables, adversarial learning, diffusion, flow, and autoregression model images?	separate reconstruction, fidelity, coverage, alignment, memorization, and sampling cost
5	Image restoration and inverse problems	Which image is supported by a corrupted or incomplete measurement, and what comes only from the prior?	specify the forward operator and separate distortion, perceptual quality, data consistency, and uncertainty
6	Video, 3D, and embodied perception	How do time, camera geometry, scene representation, and action change visual inference?	test whether a system uses motion or geometry rather than appearance shortcuts
7	Evaluation and robustness	Does a result survive shift, corruption, subgroup analysis, calibration, and real deployment constraints?	produce a model card with uncertainty, failure slices, provenance, and measured efficiency

The stages express prerequisites, but the advanced branches can be read in parallel. A detection researcher should still learn representation evaluation; a generative researcher should still understand sampling and geometry.

A common language for visual tasks

Let an image be x ∈ ℝ^(H × W × C). Different tasks change the output and therefore the valid metric:

Task	Typical output	What the output omits
classification	`p(y	x)`
detection	ranked (b_i, y_i, s_i) boxes	exact object shape
semantic segmentation	class at every pixel	object identity within one class
instance segmentation	class-labelled object masks	uncountable “stuff” unless extended to panoptic output
depth / flow	dense continuous field	semantic identity
retrieval	ranked image or text items	a calibrated decision threshold
generation	samples or a conditional distribution	proof that a requested fact is grounded
tracking	identities and states through time	complete scene understanding

Never transfer a metric merely because two systems accept images. Dice and IoU measure overlap; average precision measures ranked matching under specified thresholds; top-1 accuracy ignores localization; a perceptual generation score is not evidence of geometric correctness.

The comparison contract

For every paper or implementation, record:

1. Evidence: datasets, split unit, duplicate control, pretraining overlap, and annotations or text available at train and test time.
2. Representation: pixels, patches, regions, frames, points, masks, or learned tokens; spatial and temporal resolution; selected layer and pooling rule.
3. Learning signal: labels, paired text, augmentations, masks, teacher features, future observations, reconstruction target, or interaction.
4. Adaptation: frozen retrieval or probe, parameter-efficient tuning, or full end-to-end training.
5. Measurement: metric definition, uncertainty, slices, external shift, parameters, memory, latency, throughput, precision, and hardware.

Changing several of these at once prevents an architectural conclusion. “Foundation model” describes scope and reuse; it does not exempt a system from controlled baselines.

Suggested milestones

Milestone A: one image, three representations

On a small natural-image dataset, compare a color histogram or local descriptor, a CNN trained from scratch, and a frozen pretrained encoder with a linear head. Use one patient-free image split, one preprocessing record, and one metric script.

Milestone B: one dense task

Train either a detector or segmenter. Visualize false positives, missed small objects, merged instances, boundary errors, and confidence. Change the metric threshold without retraining to see which conclusions move.

Milestone C: one representation audit

For one self-supervised or language-aligned encoder, compare k-nearest neighbours, a frozen linear probe, low-shot tuning, and full tuning. Preserve the encoder checkpoint, input resolution, normalization, and evaluation split.

Milestone D: one dynamic or geometric test

Use temporal shuffling for video or pose perturbation for multiview reconstruction. If performance does not change, the claimed temporal or geometric mechanism may not be carrying the result.

Milestone E: a reproducible paper reconstruction

Reconstruct one paper’s input, target, objective, model path, inference procedure, metric, strongest baseline, and strongest untested assumption. Report what you verified, what you inferred, and what the released artifacts do not permit.

Bridge to computational pathology

Natural-image methods provide useful controls, not automatic solutions. Whole-slide images introduce gigapixel scale, physical magnification, stain and scanner shift, spatially sparse labels, multiple slides per case, and serious leakage risks. Carry the comparison contract into:

• WSI engineering and quality control;
• pathology self-supervision;
• foundation encoders;
• MIL and whole-slide learning;
• pathology VLMs.

The correct transfer question is not “Which natural-image model is newest?” It is “Which visual assumption, supervision source, and evaluation protocol remain valid for this WSI task?”

Dedicated spatial route

The combined video chapter introduces geometry, but camera recovery, depth, point-based learning, neural fields, Gaussian splatting, and feed-forward scene models now have a separate 3D and spatial vision route. Use it whenever coordinate frames, scale, calibration, or geometric alignment are part of the claim.