Object detection concepts in TensorFlow - ML Experiment: Train & Evaluate

Experiment - Image classification concepts

Problem:Classify objects in images correctly.

Current Metrics:Training accuracy: 98%, Validation accuracy: 65%, Training loss: 0.05, Validation loss: 0.45

Issue:The model is overfitting: training accuracy is very high but validation accuracy is low, indicating poor generalization.

Your Task

Reduce overfitting so that validation accuracy improves to at least 80% while keeping training accuracy below 90%.

You can only modify the model architecture and training hyperparameters.

Do not change the dataset or add more data.

Hint 1

Hint 2

Hint 3

Hint 4

Solution

TensorFlow

import tensorflow as tf
from tensorflow.keras import layers, models
from tensorflow.keras.preprocessing.image import ImageDataGenerator

# Data augmentation to reduce overfitting
train_datagen = ImageDataGenerator(
    rescale=1./255,
    rotation_range=20,
    width_shift_range=0.2,
    height_shift_range=0.2,
    horizontal_flip=True,
    validation_split=0.2
)

train_generator = train_datagen.flow_from_directory(
    'data/train',
    target_size=(128, 128),
    batch_size=32,
    class_mode='categorical',
    subset='training'
)

# Validation datagen without augmentation
val_datagen = ImageDataGenerator(
    rescale=1./255,
    validation_split=0.2
)

validation_generator = val_datagen.flow_from_directory(
    'data/train',
    target_size=(128, 128),
    batch_size=32,
    class_mode='categorical',
    subset='validation'
)

# Define a simpler CNN model with dropout
model = models.Sequential([
    layers.Conv2D(32, (3,3), activation='relu', input_shape=(128,128,3)),
    layers.MaxPooling2D(2,2),
    layers.Dropout(0.25),
    layers.Conv2D(64, (3,3), activation='relu'),
    layers.MaxPooling2D(2,2),
    layers.Dropout(0.25),
    layers.Flatten(),
    layers.Dense(128, activation='relu'),
    layers.Dropout(0.5),
    layers.Dense(train_generator.num_classes, activation='softmax')
])

model.compile(
    optimizer=tf.keras.optimizers.Adam(learning_rate=0.0005),
    loss='categorical_crossentropy',
    metrics=['accuracy']
)

history = model.fit(
    train_generator,
    epochs=25,
    validation_data=validation_generator
)

Added data augmentation to training images to increase variety.

Added dropout layers after convolution and dense layers to reduce overfitting.

Reduced learning rate to 0.0005 for smoother training.

Simplified model architecture by limiting number of filters and layers.

Results Interpretation

Before: Training accuracy was 98% but validation accuracy was only 65%, showing overfitting.

After: Training accuracy dropped to 88% and validation accuracy improved to 82%, showing better generalization.

Adding dropout and data augmentation helps the model learn more general features and reduces overfitting, improving validation accuracy.

Bonus Experiment

Try using transfer learning with a pre-trained model like MobileNetV2 for image classification to improve accuracy further.

💡 Hint

Freeze the base layers of the pre-trained model and train only the top layers first, then fine-tune some base layers.

Practice

1. What is the main goal of object detection in TensorFlow?

easy

A. To find and label objects in images with bounding boxes

B. To generate captions for images

C. To classify images into categories without localization

D. To enhance image resolution

Solution

1. Step 1: Understand object detection task

   Object detection locates objects in images and labels them with boxes.

2. Step 2: Compare options with definition

   Only To find and label objects in images with bounding boxes describes finding and labeling objects with boxes, matching object detection.

3. Final Answer:

   To find and label objects in images with bounding boxes -> Option A

4. Quick Check:

   Object detection = find and label objects [OK]

Hint: Object detection = find + label objects with boxes [OK]

Common Mistakes:

• Confusing object detection with image classification
• Thinking object detection generates captions
• Mixing up image enhancement with detection

2. Which part of an object detection model is responsible for extracting features from images?

easy

A. Head

B. Backbone

C. Loss function

D. Anchor generator

Solution

1. Step 1: Identify model components

   Backbone extracts features; head predicts boxes and classes.

2. Step 2: Match roles to options

   Backbone matches feature extraction role, so Backbone is correct.

3. Final Answer:

   Backbone -> Option B

4. Quick Check:

   Feature extractor = Backbone [OK]

Hint: Backbone extracts features, head predicts boxes [OK]

Common Mistakes:

• Confusing head with feature extractor
• Thinking loss function extracts features
• Mixing anchor generator with backbone

3. Given this TensorFlow code snippet for object detection model summary, what is the output shape of the backbone's last layer?

import tensorflow as tf
base_model = tf.keras.applications.MobileNetV2(input_shape=(224,224,3), include_top=False)
print(base_model.output_shape)

medium

A. (None, 7, 7, 1280)

B. (None, 14, 14, 512)

C. (None, 224, 224, 3)

D. (None, 1, 1, 1000)

Solution

1. Step 1: Understand MobileNetV2 output shape

   MobileNetV2 with include_top=False outputs feature maps of shape (None, 7, 7, 1280) for 224x224 input.

2. Step 2: Match output shape to options

   (None, 7, 7, 1280) matches expected output shape; others do not.

3. Final Answer:

   (None, 7, 7, 1280) -> Option A

4. Quick Check:

   MobileNetV2 backbone output = (None,7,7,1280) [OK]

Hint: MobileNetV2 backbone outputs 7x7x1280 feature maps [OK]

Common Mistakes:

• Confusing input shape with output shape
• Choosing output with wrong spatial dimensions
• Assuming include_top=True output shape

4. Identify the error in this TensorFlow object detection training code snippet:

model.compile(optimizer='adam', loss='categorical_crossentropy')
model.fit(train_images, train_boxes, epochs=5)

medium

A. fit method missing batch_size argument

B. Missing optimizer parameter in compile

C. Using 'categorical_crossentropy' loss instead of a detection loss

D. train_boxes should be train_labels

Solution

1. Step 1: Check loss function suitability

   Object detection needs specialized loss (e.g., localization + classification), not plain categorical_crossentropy.

2. Step 2: Validate other code parts

   Optimizer is given, train_boxes is correct for boxes, batch_size is optional.

3. Final Answer:

   Using 'categorical_crossentropy' loss instead of a detection loss -> Option C

4. Quick Check:

   Detection needs special loss, not categorical_crossentropy [OK]

Hint: Detection needs combined loss, not plain classification loss [OK]

Common Mistakes:

• Using classification loss for detection
• Confusing train_boxes with labels
• Assuming batch_size is mandatory

5. You want to improve an object detection model's accuracy on small objects. Which approach is best to try first?

hard

A. Remove the backbone and use only the head

B. Reduce number of training epochs

C. Use grayscale images instead of color

D. Increase input image resolution and adjust anchor box sizes

Solution

1. Step 1: Understand small object detection challenges

   Small objects need higher resolution and anchors matching their size for better detection.

2. Step 2: Evaluate options for improving small object accuracy

   Increasing resolution and adjusting anchors directly helps; reducing epochs or removing backbone harms performance; grayscale reduces info.

3. Final Answer:

   Increase input image resolution and adjust anchor box sizes -> Option D

4. Quick Check:

   Small objects = higher res + proper anchors [OK]

Hint: Small objects need bigger input and matching anchors [OK]

Common Mistakes:

• Reducing epochs thinking it helps accuracy
• Removing backbone reduces feature quality
• Using grayscale loses color info needed for detection