SSD concept in Computer Vision - ML Experiment: Train & Evaluate

Experiment - SSD concept

Problem:You want to detect objects in images using the Single Shot MultiBox Detector (SSD) model. The current SSD model is trained on a small dataset but shows high training accuracy (95%) and low validation accuracy (60%).

Current Metrics:Training accuracy: 95%, Validation accuracy: 60%, Training loss: 0.15, Validation loss: 0.85

Issue:The model is overfitting: it performs very well on training data but poorly on validation data.

Your Task

Reduce overfitting so that validation accuracy improves to at least 75% while keeping training accuracy below 90%.

You can only modify the model architecture and training hyperparameters.

You cannot increase the dataset size or use external data.

You must keep the SSD model framework.

Hint 1

Hint 2

Hint 3

Hint 4

Solution

Computer Vision

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

# Define a simple SSD-like model with dropout
input_shape = (300, 300, 3)
num_classes = 21  # Example for VOC dataset

inputs = layers.Input(shape=input_shape)

# Base convolutional layers
x = layers.Conv2D(32, 3, activation='relu', padding='same')(inputs)
x = layers.MaxPooling2D(2)(x)
x = layers.Conv2D(64, 3, activation='relu', padding='same')(x)
x = layers.MaxPooling2D(2)(x)

# Add dropout to reduce overfitting
x = layers.Dropout(0.3)(x)

# Additional convolutional layers
x = layers.Conv2D(128, 3, activation='relu', padding='same')(x)
x = layers.MaxPooling2D(2)(x)
x = layers.Dropout(0.3)(x)

# Prediction layers (simplified for demonstration)
x = layers.Flatten()(x)
x = layers.Dense(256, activation='relu')(x)
x = layers.Dropout(0.4)(x)
outputs = layers.Dense(num_classes, activation='softmax')(x)

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

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

# Data augmentation setup
train_datagen = ImageDataGenerator(
    rescale=1./255,
    rotation_range=15,
    width_shift_range=0.1,
    height_shift_range=0.1,
    horizontal_flip=True,
    validation_split=0.2
)

# Assuming train_dir contains training images organized in class folders
train_generator = train_datagen.flow_from_directory(
    'train_dir',
    target_size=(300, 300),
    batch_size=32,
    class_mode='categorical',
    subset='training'
)

val_datagen = ImageDataGenerator(
    rescale=1./255,
    validation_split=0.2
)

validation_generator = val_datagen.flow_from_directory(
    'train_dir',
    target_size=(300, 300),
    batch_size=32,
    class_mode='categorical',
    subset='validation'
)

# Train the model with fewer epochs
history = model.fit(
    train_generator,
    epochs=15,
    validation_data=validation_generator
)

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

Implemented data augmentation to increase data variety during training.

Lowered the learning rate from 0.001 to 0.0005 for smoother training.

Reduced the number of epochs from 50 to 15 to prevent memorization.

Results Interpretation

Before: Training accuracy 95%, Validation accuracy 60%, Training loss 0.15, Validation loss 0.85

After: Training accuracy 88%, Validation accuracy 78%, Training loss 0.30, Validation loss 0.55

Adding dropout and data augmentation helps the SSD model generalize better, reducing overfitting and improving validation accuracy.

Bonus Experiment

Try using batch normalization layers in the SSD model to improve training stability and further reduce overfitting.

💡 Hint

Insert batch normalization after convolutional layers before activation functions.

Practice

1. What is the main advantage of the SSD (Single Shot MultiBox Detector) in object detection?

easy

A. It uses multiple passes to improve accuracy.

B. It detects objects in one pass, making it faster than some other methods.

C. It only detects one object per image.

D. It requires manual feature extraction before detection.

Solution

1. Step 1: Understand SSD's detection process

   SSD predicts object locations and classes in a single forward pass through the network, unlike methods that require multiple passes.

2. Step 2: Compare with other methods

   Methods like R-CNN use multiple passes, which are slower. SSD's single pass makes it faster for real-time use.

3. Final Answer:

   It detects objects in one pass, making it faster than some other methods. -> Option B

4. Quick Check:

   SSD speed = single pass detection [OK]

Hint: SSD = Single pass detection for speed [OK]

Common Mistakes:

• Thinking SSD uses multiple passes
• Believing SSD detects only one object
• Assuming manual feature extraction is needed

2. Which of the following is the correct description of SSD's output format?

easy

A. Multiple bounding boxes with class scores predicted in one pass.

B. A single class label for the whole image.

C. A heatmap showing object locations only.

D. A sequence of images with detected objects highlighted.

Solution

1. Step 1: Recall SSD output structure

   SSD outputs many bounding boxes with associated class probabilities in one forward pass.

2. Step 2: Eliminate incorrect options

   A single class label for the whole image is wrong because SSD predicts multiple objects, not one label. A heatmap showing object locations only is incorrect as SSD outputs boxes, not just heatmaps. A sequence of images with detected objects highlighted describes video or image sequence output, not SSD's prediction format.

3. Final Answer:

   Multiple bounding boxes with class scores predicted in one pass. -> Option A

4. Quick Check:

   SSD output = boxes + class scores [OK]

Hint: SSD outputs many boxes with classes at once [OK]

Common Mistakes:

• Confusing SSD output with single-label classifiers
• Thinking SSD outputs only heatmaps
• Assuming SSD outputs image sequences

3. Consider this simplified SSD output for one image:
boxes = [[0.1, 0.1, 0.4, 0.4], [0.5, 0.5, 0.9, 0.9]]
scores = [[0.9, 0.1], [0.2, 0.8]]
Here, each score list corresponds to [class0, class1]. Which box is predicted as class1 with higher confidence?

medium

A. The first box [0.1, 0.1, 0.4, 0.4]

B. Neither box predicts class1

C. Both boxes have equal confidence for class1

D. The second box [0.5, 0.5, 0.9, 0.9]

Solution

1. Step 1: Identify class1 scores for each box

   First box class1 score = 0.1, second box class1 score = 0.8.

2. Step 2: Compare scores to find higher confidence

   0.8 > 0.1, so second box has higher confidence for class1.

3. Final Answer:

   The second box [0.5, 0.5, 0.9, 0.9] -> Option D

4. Quick Check:

   Class1 highest score = 0.8 at second box [OK]

Hint: Check class1 scores, pick highest [OK]

Common Mistakes:

• Mixing class0 and class1 scores
• Ignoring score values and guessing
• Assuming equal confidence without checking

4. You have an SSD model that predicts many overlapping boxes for the same object. What common step should you add to fix this issue?

medium

A. Increase the learning rate during training.

B. Reduce the number of classes predicted.

C. Apply Non-Maximum Suppression (NMS) to remove overlapping boxes.

D. Use grayscale images instead of color.

Solution

1. Step 1: Understand overlapping box problem

   SSD often predicts multiple boxes around the same object, causing duplicates.

2. Step 2: Identify solution to remove duplicates

   Non-Maximum Suppression (NMS) filters overlapping boxes by keeping the one with highest confidence and removing others.

3. Final Answer:

   Apply Non-Maximum Suppression (NMS) to remove overlapping boxes. -> Option C

4. Quick Check:

   Overlapping boxes fix = NMS [OK]

Hint: Use NMS to clean overlapping boxes [OK]

Common Mistakes:

• Changing learning rate won't fix overlaps
• Reducing classes doesn't remove duplicates
• Switching image color doesn't affect box overlap

5. In a real-time traffic camera system using SSD, which combination of factors best balances speed and accuracy?

hard

A. Use SSD with a lightweight backbone network and moderate input image size.

B. Use SSD with a very deep backbone and very large input images.

C. Use SSD with no pre-trained weights and very small input images.

D. Use SSD with multiple passes over the image for better accuracy.

Solution

1. Step 1: Consider real-time constraints

   Real-time systems need fast processing, so a lightweight backbone helps speed up SSD.

2. Step 2: Balance input size and accuracy

   Moderate image size keeps enough detail for accuracy without slowing down processing too much.

3. Step 3: Evaluate other options

   Very deep backbones and large images slow down speed (Use SSD with a very deep backbone and very large input images.). No pre-training and very small images reduce accuracy (Use SSD with no pre-trained weights and very small input images.). Multiple passes slow down speed (Use SSD with multiple passes over the image for better accuracy.).

4. Final Answer:

   Use SSD with a lightweight backbone network and moderate input image size. -> Option A

5. Quick Check:

   Speed + accuracy = lightweight backbone + moderate size [OK]

Hint: Light backbone + moderate size = best speed-accuracy [OK]

Common Mistakes:

• Choosing very deep networks for real-time
• Using very small images losing accuracy
• Running multiple passes slows system