monocular-depth-estimation-keras

Depth estimation from a single image is a challenging yet fascinating problem in computer vision. From my experience as a Python Keras developer, building a monocular depth estimation model can unlock many applications like robotics, AR, and autonomous navigation.

In this tutorial, I’ll show you how to create a depth estimation model using Keras. I’ll provide complete code snippets for each step, making it easy for you to follow and apply.

Table of Contents

What is Monocular Depth Estimation in Python Keras?

Monocular depth estimation predicts the distance of objects from a single RGB image, unlike stereo methods that rely on two cameras. Using Keras, you can build convolutional encoder-decoder networks that learn to infer depth maps.

This approach is practical and widely used in real-world applications where only one camera is available.

Method 1: Build a Simple Encoder-Decoder Depth Estimation Model in Keras

This method uses a convolutional encoder-decoder architecture to predict depth maps from input images.

Step 1: Import Required Libraries

Import the core TensorFlow and helper libraries needed to build the model.

import tensorflow as tf
from tensorflow.keras import layers, models
import numpy as np

Step 2: Define the Encoder-Decoder Model

Define a simple encoder–decoder network that maps images to depth predictions.

def create_depth_estimation_model(input_shape=(128, 128, 3)):
    inputs = layers.Input(shape=input_shape)

    # Encoder: Convolutional layers to extract features
    x = layers.Conv2D(64, 3, activation='relu', padding='same')(inputs)
    x = layers.MaxPooling2D(2)(x)
    x = layers.Conv2D(128, 3, activation='relu', padding='same')(x)
    x = layers.MaxPooling2D(2)(x)
    x = layers.Conv2D(256, 3, activation='relu', padding='same')(x)
    x = layers.MaxPooling2D(2)(x)

    # Decoder: Upsampling layers to reconstruct depth map
    x = layers.Conv2DTranspose(256, 3, strides=2, activation='relu', padding='same')(x)
    x = layers.Conv2DTranspose(128, 3, strides=2, activation='relu', padding='same')(x)
    x = layers.Conv2DTranspose(64, 3, strides=2, activation='relu', padding='same')(x)

    # Output depth map with single channel (grayscale)
    outputs = layers.Conv2D(1, 3, activation='sigmoid', padding='same')(x)

    model = models.Model(inputs, outputs)
    return model

model = create_depth_estimation_model()
model.summary()

Step 3: Compile the Model

Compile the model using an optimizer and loss function suitable for depth estimation.

model.compile(optimizer='adam', loss='mse', metrics=['mae'])

Step 4: Prepare Dummy Data for Training

Create sample training and validation data to demonstrate how the model works.

# Generate 100 dummy RGB images (128x128x3)
X_train = np.random.rand(100, 128, 128, 3).astype(np.float32)
# Generate 100 dummy depth maps (128x128x1)
y_train = np.random.rand(100, 128, 128, 1).astype(np.float32)

X_val = np.random.rand(20, 128, 128, 3).astype(np.float32)
y_val = np.random.rand(20, 128, 128, 1).astype(np.float32)

Step 5: Train the Model

Train the depth estimation model on the prepared dataset for learning.

history = model.fit(
    X_train, y_train,
    validation_data=(X_val, y_val),
    epochs=15,
    batch_size=8
)

You can refer to the screenshot below to see the output.

Method 2: Use a Pretrained Encoder for Better Feature Extraction

Using a pretrained backbone like MobileNetV2 improves feature extraction and speeds up convergence.

Step 1: Load Pretrained MobileNetV2 as Encoder

In this step, we load MobileNetV2 as a pretrained encoder to extract high-level image features, giving the depth-estimation model a stronger starting point and reducing the training effort.

def create_depth_model_with_mobilenet(input_shape=(128, 128, 3)):
    base_model = tf.keras.applications.MobileNetV2(
        input_shape=input_shape,
        include_top=False,
        weights='imagenet'
    )

    # Freeze encoder weights
    base_model.trainable = False

    inputs = layers.Input(shape=input_shape)
    x = base_model(inputs, training=False)

    # Decoder: Upsampling to reconstruct depth map
    x = layers.Conv2DTranspose(256, 3, strides=2, activation='relu', padding='same')(x)
    x = layers.Conv2DTranspose(128, 3, strides=2, activation='relu', padding='same')(x)
    x = layers.Conv2DTranspose(64, 3, strides=2, activation='relu', padding='same')(x)
    x = layers.Conv2DTranspose(32, 3, strides=2, activation='relu', padding='same')(x)

    outputs = layers.Conv2D(1, 3, activation='sigmoid', padding='same')(x)

    model = models.Model(inputs, outputs)
    return model

model = create_depth_model_with_mobilenet()
model.summary()

Step 2: Compile and Train

After building the model, we compile it with an appropriate optimizer and loss function, then train it using our dataset to learn the mapping from RGB images to depth maps.

model.compile(optimizer='adam', loss='mse', metrics=['mae'])

model.fit(
    X_train, y_train,
    validation_data=(X_val, y_val),
    epochs=15,
    batch_size=8
)

You can refer to the screenshot below to see the output.

Tips for Improving Monocular Depth Estimation in Keras

Use data augmentation like random flips and brightness changes to improve robustness.
Experiment with different loss functions, such as scale-invariant loss, for better depth accuracy.
Fine-tune pretrained encoders after initial training for improved performance.

Monocular depth estimation in Keras is achievable with easy encoder-decoder models or by leveraging pretrained backbones. The examples I shared provide a solid foundation for your projects.

Feel free to adapt the architectures to your dataset and experiment with training strategies for the best results.

Other Python Keras articles you may also like:

Bijay Kumar

I am Bijay Kumar, a Microsoft MVP in SharePoint. Apart from SharePoint, I started working on Python, Machine learning, and artificial intelligence for the last 5 years. During this time I got expertise in various Python libraries also like Tkinter, Pandas, NumPy, Turtle, Django, Matplotlib, Tensorflow, Scipy, Scikit-Learn, etc… for various clients in the United States, Canada, the United Kingdom, Australia, New Zealand, etc. Check out my profile.

enjoysharepoint.com/

Monocular Depth Estimation Using Keras