wenmin-wu/cv-stroke-temporal-color-rendering

Stroke Temporal Color Rendering

Overview

Hand-drawn sketches are stored as ordered stroke sequences. When rendering to an image, encoding the drawing order as pixel intensity — earlier strokes brighter, later strokes darker — gives CNNs an additional temporal signal beyond pure geometry. This simple trick improves doodle classification accuracy by 1-3% over flat-color rendering.

Quick Start

import cv2
import numpy as np
import json

def draw_strokes(raw_strokes, size=64, line_width=6):
    img = np.zeros((256, 256), np.uint8)
    for t, stroke in enumerate(raw_strokes):
        color = max(255 - t * 13, 30)  # fade from 255 to ~30
        for i in range(len(stroke[0]) - 1):
            cv2.line(img,
                     (stroke[0][i], stroke[1][i]),
                     (stroke[0][i + 1], stroke[1][i + 1]),
                     color, line_width)
    if size != 256:
        img = cv2.resize(img, (size, size))
    return img

strokes = json.loads(drawing_string)
img = draw_strokes(strokes, size=64)

Workflow

1. Parse stroke JSON: list of [[x0, x1, ...], [y0, y1, ...]] per stroke
2. Create a blank canvas (256x256 grayscale)
3. Draw each stroke with intensity 255 - t * step, clamped to a minimum
4. Resize to target CNN input size (64x64 or 128x128)
5. Normalize to [0, 1] or apply model-specific preprocessing

Key Decisions

• Fade rate: 13 per stroke works for ~20-stroke drawings; adjust for longer sequences
• Minimum intensity: clamp at 30 to keep late strokes visible
• Line width: 6px at 256x256 base; scale proportionally for smaller canvases
• vs. flat color: temporal encoding adds ~1-3% top-3 accuracy at zero computational cost
• vs. RGB channels: encode first/middle/last strokes in R/G/B for richer signal with pretrained models

References

• Greyscale MobileNet [LB=0.892]