Spaces:
Runtime error
Runtime error
Commit History
-
docs: clarify COCO is pretrained-weights-only, not a training dataset
Remove the COCO row from the "Datasets at a glance" table so it lists only the two datasets actually trained on (Stanford Cars, CarDD). Add notes: CarDD ships in COCO *format* (not the COCO dataset), and the car gate uses torchvision's COCO-pretrained Mask R-CNN at inference (weights only — nothing trains on COCO). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>0f12cd5 -
docs: rework architecture explainer for the checkpoint-16 build
- Scope to what this branch actually runs: Mask R-CNN is the car GATE only (COCO, inference); damage is Variant A (ResNet-50 classifier) + Variant B (YOLOv8). Removed the damage-segmenter content; added a clarifying note and a YOLOv8 detector diagram + a YOLO multi-part loss panel. - Fix nav: scroll-padding-top so anchor links land below the sticky nav. - Embed 6 real dataset thumbnails (Stanford Cars + CarDD, base64, self-contained). - Add a "Datasets at a glance" table and a full Citations section crediting Stanford Cars (Krause 2013), CarDD (Wang 2023), COCO, Mask R-CNN, ResNet, YOLOv8/Ultralytics, XGBoost, PyTorch. - Animate the pipeline (flowing packets + pulsing stages) and the output (annotation boxes draw in sequence). Verified in-browser: 10 sections, 6 images loaded, all canvases painting, anchors resolve below the nav. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>8e2ecca
-
docs: canvas-driven neural-network architecture explainer
Self-contained HTML (docs/network-architecture.html, no build step) explaining the ccdp models with live <canvas> diagrams: end-to-end code flow, data parsing (Stanford Cars .mat + CarDD COCO), ResNet-50 / Mask R-CNN-FPN architecture, the activation functions (ReLU/Sigmoid/Softmax/SiLU), two-stage training + augmentation, backprop/gradient flow, the per-task loss functions, accuracy metrics with the repo's production numbers, and the annotated-image output. Verified rendering: 9 sections, 16 canvases all painting. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>739fb85