wenmin-wu/cv-cnn-rnn-video-classification

CNN-RNN Video Classification

Overview

For video classification that benefits from temporal context (action recognition, deepfake detection), extract fixed-length feature vectors from each frame using a pretrained CNN (InceptionV3, ResNet), then feed the sequence to a GRU/LSTM. A boolean mask handles variable-length videos by ignoring padded positions. This two-stage approach decouples spatial feature learning from temporal modeling.

Quick Start

import tensorflow as tf
from tensorflow import keras

feature_extractor = keras.applications.InceptionV3(
    weights="imagenet", include_top=False, pooling="avg")

MAX_SEQ = 30
FEAT_DIM = 2048

frame_input = keras.Input((MAX_SEQ, FEAT_DIM))
mask_input = keras.Input((MAX_SEQ,), dtype="bool")
x = keras.layers.GRU(16, return_sequences=True)(frame_input, mask=mask_input)
x = keras.layers.GRU(8)(x)
x = keras.layers.Dropout(0.4)(x)
x = keras.layers.Dense(8, activation="relu")(x)
output = keras.layers.Dense(1, activation="sigmoid")(x)

model = keras.Model([frame_input, mask_input], output)
model.compile(loss="binary_crossentropy", optimizer="adam")

Workflow

1. Sample frames from each video (fixed max, e.g., 30)
2. Extract per-frame features: feature_extractor.predict(frame) → (2048,) vector
3. Stack into a matrix of shape (max_seq, feat_dim), zero-pad shorter videos
4. Create a boolean mask: True for real frames, False for padding
5. Train GRU on (features, mask) → binary label
6. At inference, apply the same sampling and feature extraction

Key Decisions

• CNN backbone: InceptionV3 (2048-d) or EfficientNet (1280-d); freeze weights for speed
• RNN type: GRU is faster than LSTM with comparable performance for short sequences
• Stacking: 2 GRU layers (16→8 units) is sufficient; deeper stacks overfit on small datasets
• Masking: essential for variable-length inputs — without it, the GRU learns to predict from padding
• vs. 3D CNN: CNN-RNN is more parameter-efficient and easier to pretrain; 3D CNNs capture fine-grained motion better

References

• Deep Fake Detection on Images and Videos