Single Object Localization

A PyTorch implementation of single-object localization using the Pascal VOC 2012 dataset.

The model is trained to determine whether a target object is present in an image and, when present, predict its bounding box.

For this implementation, the target object is horse.

---

Overview

This project focuses on single-object localization.

For each input image, the model produces two outputs:

1. Classification — determines whether the target object is present.
2. Bounding box regression — predicts the location of the target object.

The bounding box is represented as:

(xmin, ymin, xmax, ymax)

Architecture

                    Input Image
                         │
                         ▼
                MobileNetV3-Large
                         │
                         ▼
                  Feature Extractor
                         │
                    960 × 7 × 7
                         │
                       Flatten
                         │
                    47,040 features
                         │
              ┌──────────┴──────────┐
              │                     │
              ▼                     ▼
       Classification        BBox Regression
            Head                   Head
              │                     │
              ▼                     ▼
          1 logit             4 coordinates
                              (x1,y1,x2,y2)

The MobileNetV3-Large feature extractor produces a spatial feature map of size 960 × 7 × 7 for a 224 × 224 input. This feature map is flattened before being passed to the two prediction heads.

---

Dataset

The project uses the Pascal VOC 2012 dataset.

For this implementation, the target category is horse.

Images are filtered so that they contain at most one instance of the target object.

The resulting dataset contains:

Dataset	Images
Original dataset	5,717
Filtered dataset	5,635
Training set	4,508
Validation set	1,127

The training set contains:

• 129 positive images containing a horse
• 4,379 negative images without a horse

Since the dataset is highly imbalanced, a balanced sampler is used during training with a 3:1 negative-to-positive ratio.

---

Preprocessing

The input images are resized and center-cropped to:

224 × 224

The bounding boxes undergo the same spatial transformations as the input images so that they remain aligned with the transformed images.

Bounding-box coordinates are then normalized to the range [0, 1].

Each bounding box is represented as:

(xmin, ymin, xmax, ymax)

For negative images, where no horse is present, the bounding box is represented as:

[0, 0, 0, 0]

The images are also normalized using the standard ImageNet normalization used by the pretrained MobileNetV3-Large weights.

---

Loss Function

The model uses two loss functions.

Classification Loss

Binary cross entropy with logits is used for the classification task:

BCEWithLogitsLoss

Bounding Box Loss

Smooth L1 loss is used for bounding box regression:

SmoothL1Loss

The bounding box loss is calculated only for positive images.

The total loss is:

Total Loss = Classification Loss + Bounding Box Loss

---

Training

The model is trained using:

Parameter	Value
Backbone	MobileNetV3-Large
Optimizer	Adam
Learning rate	1e-4
Batch size	32
Epochs	10
Negative-to-positive ratio	3:1

The MobileNetV3-Large backbone is initialized with pretrained ImageNet weights.

The model is trained end-to-end, with separate heads for classification and bounding box regression.

The best model checkpoint is selected based on validation Mean IoU.

---

Evaluation

The model is evaluated using both classification and localization metrics.

Classification Metrics

• Accuracy
• Precision
• Recall
• F1 Score
• Confusion Matrix

Localization Metrics

• Mean IoU
• Percentage of predictions with IoU ≥ 0.50
• Percentage of predictions with IoU ≥ 0.75

Intersection over Union (IoU) measures the overlap between the predicted bounding box and the ground-truth bounding box.

---

Results

The final model achieves the following results on the validation set.

Classification

Metric	Result
Accuracy	98.85%
Precision	70.59%
Recall	88.89%
F1 Score	78.69%

Confusion Matrix

	Predicted Positive	Predicted Negative
Actual Positive	24	3
Actual Negative	10	1,090

Localization

Metric	Result
Mean IoU	0.6042
IoU ≥ 0.50	66.67%
IoU ≥ 0.75	18.52%

---

Visualization

The project includes a visualization script for inspecting the predicted bounding boxes.

For positive validation images, the visualization displays:

• Ground-truth bounding box
• Predicted bounding box
• Horse classification probability
• IoU between the predicted and ground-truth bounding boxes

Run:

python src\visualize.py

This allows the localization performance to be inspected visually in addition to the numerical evaluation metrics.

---

Project Structure

SingleObjectLocalization/
│
├── .data/
│   └── VOCdevkit/
│
├── src/
│   ├── dataset.py
│   ├── transforms.py
│   ├── model.py
│   ├── loss.py
│   ├── train.py
│   ├── evaluate.py
│   └── visualize.py
│
├── notebooks/
│
├── data/
├── checkpoints/
├── outputs/
│
├── README.md
├── requirements.txt
├── .gitignore
└── single_object_model.pth

The dataset, generated outputs, checkpoints, and model weights are excluded from version control through .gitignore.

---

Installation

Clone the repository:

git clone <repository-url>
cd SingleObjectLocalization

Create a virtual environment:

python -m venv .venv

Activate the virtual environment on Windows:

.venv\Scripts\Activate.ps1

Install the required dependencies:

pip install -r requirements.txt

---

Dataset Setup

Download the Pascal VOC 2012 dataset.

Place the dataset under:

.data/

The expected structure is:

.data/
└── VOCdevkit/
    └── VOC2012/
        ├── Annotations/
        ├── ImageSets/
        ├── JPEGImages/
        ├── SegmentationClass/
        └── SegmentationObject/

The dataset is not included in the repository because of its size.

---

Training

From the project root, run:

python src\train.py

The training script:

1. Loads the Pascal VOC dataset.
2. Filters images based on the target category.
3. Applies image and bounding-box preprocessing.
4. Creates balanced training batches.
5. Trains the MobileNetV3-based localization model.
6. Evaluates the model on the validation set after each epoch.
7. Saves the checkpoint with the best validation Mean IoU.

The best model is saved as:

single_object_model.pth

---

Evaluation

After training, evaluate the saved model using:

python src\evaluate.py

This reports:

• Accuracy
• Precision
• Recall
• F1 Score
• Confusion Matrix
• Mean IoU
• IoU ≥ 0.50
• IoU ≥ 0.75

---

Visualization

To visualize predictions on validation images:

python src\visualize.py

The script loads the saved model and displays examples containing the target object.

---

Requirements

The main dependencies are:

• Python
• PyTorch
• TorchVision
• NumPy
• Matplotlib
• Pillow
• tqdm

The exact installed versions are specified in:

requirements.txt

---

Notes

Spatial Feature Extraction

The MobileNetV3-Large classification head is not used.

Instead, the model uses the spatial feature extractor:

mobilenet_v3_large(weights=weights).features

For a 224 × 224 input, this produces:

960 × 7 × 7

The resulting feature map is flattened:

960 × 7 × 7 = 47,040

These features are then passed independently to the classification and bounding-box regression heads.

Bounding Box Prediction

The bounding-box head outputs four normalized coordinates:

(xmin, ymin, xmax, ymax)

A sigmoid activation keeps the predicted coordinates within the [0, 1] range.

Negative Images

For images without a horse:

label = 0
bbox = [0, 0, 0, 0]

The bounding-box loss is not applied to these negative samples.

---

Future Work

This project focuses specifically on single-object localization.

Possible extensions include:

• Multi-object detection
• Multiple bounding boxes per image
• Different target categories
• More advanced detection architectures
• Non-Maximum Suppression
• Mean Average Precision (mAP)
• Faster R-CNN
• YOLO-style object detection