Contact person: Business Client

Phone:Show

Email:Show

Location: Altenfelden, Austria

Budget: Recommended by industry experts

Time to start: As soon as possible

Project description:
"Build a production quality WPF desktop application (Visual Studio 2022, .NET Framework 4.7.2, C# 7.3) that removes photographic backgrounds and composites people (including held objects / props) onto virtual backgrounds. The deliverable must preserve fine hair detail, handle varied image inputs and lighting, and produce an “organic” final composite.

Key goals & constraints

• Best possible visual quality across diverse samples: phone, webcam and DSLR images; indoor/outdoor and cluttered backgrounds; subjects holding objects.
• Use models that are permissively licensed for commercial use where possible. Preferred examples: U2 Net, MODNet (contractor must confirm license and document suitability). Do not rely on paid/closed APIs (e.g., rmbg).
• Use ONNX for inference and integrate with Microsoft ML tooling ([login to view URL] / [login to view URL] support or ONNX Runtime usage). OpenCvSharp (or other free imaging libraries) is acceptable for preprocessing/postprocessing.
• GPU acceleration is optional - implement ONNX Runtime providers (DirectML / CUDA) where feasible but provide a robust CPU fallback. Expect many target machines will be CPU only.
• Keep final code compatible with Visual Studio 2022 and the existing WPF project ([login to view URL] / [login to view URL]). Provide solutions compatible with .NET Framework 4.7.2 or explain necessary migration steps.

Functional requirements

• Input: load image(s) and virtual background(s) from disk (file dialog will do).
• Output: composited image (subject on chosen background) with transparent/antialiased edges and preserved hair detail and held objects. Support saving as PNG and JPEG.
• Processing pipeline: optional resize/scale to MaxWidth/MaxHeight, run ONNX model to generate mask(s), refine mask (morphology, erosion, inner solid mask, feathering, blur), composite with background, optional underlay for thin hair.
• Expose runtime parameters (sliders / UI controls) similar to current app: MaxWidth, MaxHeight, ConfidenceThreshold, FeatherLow/High, ErodeSize, InnerSolidSize, BlurKernel, EnableWhiteUnderlay.
• Provide automated bench/test runner to measure processing times on supplied samples.

Non functional & performance targets

• Full HD (1920x1080) composited result should preferably process in under 1 second on a modern CPU.
• 24MP image target: aim ~3 seconds on a 6th gen i5 (2 cores / 4 threads, 8GB RAM) — realistic best effort; document actual measured timings and test hardware.
• Implementation must be thread safe for UI responsiveness (use [login to view URL] / background worker for processing).
• Deliver clean, maintainable C# code, documented public APIs, and small set of unit/integration tests where applicable.

Technical requirements & recommendations

• Use ONNX model (U2 Net / MODNet or equivalent). Contractor must:
o Propose model(s) and confirm commercial license compatibility.
o Provide a short justification for chosen model with expected tradeoffs (quality vs. speed).
• Use ONNX Runtime (with optional DirectML/CUDA providers) or [login to view URL]'s ONNX support. Must include CPU fallback.
• Use OpenCvSharp for image resizing, blurring, morphology, and precise mask manipulation.
• Implement mask refinement steps: confidence thresholding, erosion for inner-solid mask, Gaussian blur / feathering, and compositing with alpha blending that preserves hair transparency.
• Provide instructions for installing any native dependencies and steps to enable optional GPU acceleration.
• The delivered solution should integrate easily with the existing WPF MainWindow + BackgroundRemover patterns (no unnecessary project rearchitecture).
Deliverables
• Full Visual Studio 2022 solution (.sln) targeting .NET Framework 4.7.2.
• Source code for BackgroundRemover with clean public API, implemented inference pipeline, and mask refinement logic.
• Small WPF demo app (improved Version you get from me) that reproduces UI flow: load image/background, preview, apply, save. Preserve existing UI parameters.
• A set of test/sample images and a benchmarking script that outputs average/min/max process times.
• Short video or GIF (optional) showing before/after on at least 3 diverse samples.

Acceptance criteria

• Demonstrated preservation of fine hair detail and good separation of people + held objects across at least 10 diverse sample images.
• Measured performance data attached; for full HD images processing time ≤ 1s in the test environment claimed by the contractor, and reasonable timings for the 24MP target with documented hardware.
• Source builds cleanly in Visual Studio 2022; solution opens and runs without extra paid dependencies.
• License verification included for models and libraries, with explicit statement they can be used commercially (or an alternative if not).

Proposal requirements (what I ask from bidders)

• Short technical approach (model choice, ONNX runtime provider, mask refinement algorithm).
• Confirmation of commercial license for the chosen model(s) and link to license text.
• Example screenshots or short video of prior similar work.
• Expected delivery items and timeline (timeline flexible; quality is primary).
• Fixed price and any optional items (GPU support, additional postprocessing).
• Clarify any deviations from .NET Framework 4.7.2 / WPF requirement up front.

Notes for bidders

• Preference to contractors who have experience optimizing ONNX models for CPU inference and who can demonstrate handling of hair/edge cases.
• Avoid paid third party APIs; the solution should run locally and be deployable.

What you get from me to start the Project?

• An already (kind of) working example of the Demo application with various parameters already implemented to test
• A selection of different Images which should be processed with the application." (client-provided description)