| name | cv-conv1d-lstm-stroke-classifier |
| description | Stack 1D convolutions for local feature extraction before bidirectional LSTMs to classify variable-length stroke sequences into hundreds of doodle categories |
Conv1D-LSTM Stroke Classifier
Overview
Sketches can be represented as sequences of (x, y, stroke_id) points rather than rasterized images. A Conv1D-LSTM architecture processes this raw sequence: 1D convolutions extract local patterns (corners, curves), then LSTMs capture long-range stroke order dependencies. This approach works directly on stroke coordinates without rendering, preserving resolution-independent information.
Quick Start
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import (Conv1D, LSTM, Dense,
Dropout, BatchNormalization)
n_classes = 340
max_points = 128
model = Sequential([
BatchNormalization(input_shape=(max_points, 3)),
Conv1D(48, 5, activation="relu"),
Dropout(0.3),
Conv1D(64, 5, activation="relu"),
Dropout(0.3),
Conv1D(96, 3, activation="relu"),
Dropout(0.3),
LSTM(128, return_sequences=True),
Dropout(0.3),
LSTM(128, return_sequences=False),
Dropout(0.3),
Dense(512, activation="relu"),
Dropout(0.3),
Dense(n_classes, activation="softmax"),
])
model.compile(optimizer="adam", loss="categorical_crossentropy",
metrics=["accuracy"])
Workflow
- Parse strokes into Nx3 matrix: (x, y, stroke_flag) where flag marks stroke boundaries
- Pad or truncate to fixed length (128-256 points)
- Normalize x, y coordinates to [0, 1]
- Feed through Conv1D layers (increasing filters: 48→64→96) for local pattern extraction
- Feed through stacked LSTMs for sequential modeling
- Dense layers → softmax over N classes
Key Decisions
- Input format: (x, y, stroke_flag) preserves drawing order; stroke_flag = 1 (continue) or 2 (new stroke)
- Sequence length: 128 captures most sketches; longer sequences add diminishing returns
- Conv before LSTM: Conv1D reduces sequence length and extracts local features, making LSTM more effective
- Dropout: 0.3 between every layer prevents overfitting on the large number of classes
- vs. image CNN: stroke-based models are resolution-independent and faster; image CNNs are more accurate with pretrained weights
References