Week 6 Progress Report – Final Touches & Presentation Preparation #

1 Executive Summary #

Week 6 concluded the implementation phase and opened the release-preparation window.
The core product functionality—automatic tyre analysis via tread-depth regression, spike-condition segmentation, and brand-text OCR—has remained stable since the Week 5 freeze candidate. This final sprint therefore emphasised four themes:

1. Performance & Portability – critical models were exported to ONNX, yielding up-to-10× CPU speed-ups (2 s → 0.3 s per inference) and lowering memory overhead;
2. UX Polish – the Web application gained quick-action buttons, friendlier waiting dialogues, local chat-history persistence, and was re-deployed to Vercel for improved availability;
3. Synthetic-Data Pipeline Kick-off – a Blender workflow capable of rendering tyres with programmatically varied tread depth produced its first sample set, seeding future augmentation;
4. Proxy Integration – All outgoing requests are routed through TOR proxy network to avoid country-level blocks

The system now meets functional requirements and is locked for presentation, with only critical hot-fixes allowed until final defence.

2 Final-Week Technical Work #

2.1 Model Optimisation & Conversion #

Conversion process. All exports use opset 17, validated with onnx-runtime 1.17.1. Batch-norm folding and constant-fold passes were applied via onnx-optimizer; the resulting artefacts live under artifacts/onnx/. Automated integrity tests confirm numerical parity (<1e-5 MSE) with the reference PyTorch graphs.

2.2 Web-Site Enhancements #

Quick-action buttons. The landing page now offers “Analyse Tread”, “Analyse Spike”, and “Detect Brand” actions that route the uploaded image straight to the corresponding API without an intermediary menu click, trimming the median user pathway by one interaction.

Waiting dialogue. A progressive-disclosure modal supplies status (“uploading”, “pre-processing”, “inference”) with time-outs and fallback guidance. This reduced observed early-abandon events during internal tests.

Local-storage chat history. User conversations (image thumbnails plus textual results) persist in window.localStorage. A newly-added toolbar icon allows clearing the cache, satisfying privacy feedback gathered in Week 5.

Deployment. The static React bundle and serverless proxy were re-deployed to Vercel (region fra1).

2.3 New Unwrapper Model & Dataset Expansion #

The earlier open-source tyre-ring detector lacked robustness when wheels were mounted or background clutter appeared. This week’s replacement:

• Architecture – SegFormer-B2 backbone with mixed-scale deformable-attention decoder;
• Training corpus – 1 260 pre-labelled “bare” tyres plus 200 new wheel-mounted instances (manual correction after auto-label seeding);
• Augmentation – random perspective warp, synthetic background compositing, Gaussian-noise injection;
• Metrics – mIoU 91 % (old model 86 %);
• Inference – ONNX FP32 on CPU 0.45 s median.

The improved mask regularity produces tighter ellipse-fitting, which in turn sharpens ROI cropping for downstream OCR.

2.4 Blender Variable-Depth Tyre Model #

Pipeline overview.

1. Vectorise a 2-D tread-pattern image;
2. Extrude along normal to form height field;
3. Parameterise depth δ ∈ [1 mm, 10 mm];
4. Render colour pass (1024 × 1024) plus Z-buffer;
5. Export depth in millimetres via normalised EXR.

Current status. First tyre instance rendered; 50 images covering five depth settings exported. Lighting and background realism remain work-in-progress; glossy floor reflections are being tuned via Cycles node-graph to minimise domain shift. The initial renders will seed a pilot experiment on depth-regression fine-tuning next sprint.

3 Governance & Release Management #

3.1 Testing & Coverage (post-freeze) #

• Backend unit tests: 9 / 9 pass – 100 % statement coverage.
• API integration: 6 scenarios pass.
• Front-end Cypress: 34 tests pass; Lighthouse accessibility score 95 / 100.
• Telegram-bot scripted flows: 3/3 pass.

4 TOR integration #

• Added open-source dperson/torproxy container to docker-compose as a proxy
• Enabled routing all outgoing requests through proxy

Result: all VLM models from openrouter.ai are now available: no country-level blockages

5 Presentation Preparation #

Slide deck. Skeleton outline prepared (nine main slides, four backup), placeholders for screenshots awaiting final polish.

Speaker allocation. Unchanged from Week 5; rehearsal schedule booked for 25 July afternoon, 27 July morning.

Demo logistics. Live demo served from Vercel front-end with back-end API on staging VDS; fallback screencast recorded in MP4.

7 Team Contributions #

8 Project Readiness Checklist #

The project is therefore ready for final rehearsal, stakeholder review, and subsequent public defence. All further commits between the freeze and presentation will be confined to emergency patches, ensuring stability for evaluators and external testers.