[barhanc]

Description

This PR introduces new depth estimation task implementation along with a test suite and a simple example screen in computer vision app to test the new feature end-to-end. Alongside it we upload 3 depth estimation models to SWM HuggingFace repo to use with this new task with different performance/quality tradeoffs; all accessible in the new collection https://huggingface.co/collections/software-mansion/depth-estimation.

Introduces a breaking change?

• Yes
• No

Type of change

• Bug fix (change which fixes an issue)
• New feature (change which adds functionality)
• Documentation update (improves or adds clarity to existing documentation)
• Other (chores, tests, code style improvements etc.)

Tested on

• iOS
• Android

Testing instructions

• Run new example screen on device (both Android and iOS) and test newly added models.
• Check the HF repos from the linked collection https://huggingface.co/collections/software-mansion/depth-estimation,

Screenshots

Related issues

Closes #1372

Checklist

• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have updated the documentation accordingly
• My changes generate no new warnings

Additional notes

[msluszniak]

HF configs fail the spec schema:

1. family: depth-anything-v2/-v3 don't match ^[a-z0-9_]+$, should be depth_anything_v2/_v3.
2. capabilities: depth-estimation isn't in the enum. Neither are gemma-4-multimodal's audio-text-to-text/image-text-to-text, so let's add all three to the spec.
3. Variants carry quantized/default, which the schema disallows (also in the blazeface and rfdetr-keypoint Vulkan configs).

Cards:
4. ZipDepth's "Repository structure" has no sizes on the config.json lines.
5. ZipDepth Core ML is fp32, while other CV Core ML builds are fp16 (fp32 never reaches the ANE). If fp16 hurts accuracy, please note that on the card.
6. Please add an "XNNPACK int8" section with quality numbers to both Depth Anything cards, like the other int8 repos have.

[msluszniak]

Code suggestions:

1. download-libs.js: the FEATURE_MAP comment names only Depth Anything V2; V3 and ZipDepth ship the same three backends.
2. Docs still describe only V2:
   • Intro: add ZipDepth (inverse depth) and DA3 (depth, larger is farther).
   • Options table: the "Depth Anything V2 preset" column doesn't fit ZipDepth ('linear', { alpha: 1 / 255, beta: 0 }).
   • IMAGENET_NORM is used by V3 too.
3. Demo: V3 outputs depth, and 1 - normalized squashes near-range contrast (the vase merges with the table). Consider inverting before normalizing:

   const values = isDepth ? depth.data.map((d) => (d > 0 ? 1 / d : NaN)) : depth.data;

4. Demo: use the option's family instead of Object.values(DEPTH_V3_MODELS).includes(selectedModel).

Optional follow-up: a depthKind: 'inverse' | 'depth' field on DepthEstimatorModel would save apps from special-casing V3.

[msluszniak]

Tested both iOS and Android in demo apps on all possible models and worked really cool!

[barhanc]

FIxed both the HF and applied the code suggestion. Also improved the performance of models after profiling (around 10% inference time improvement).