Car Viewpoint Classification

Project Overview - About project

This project classifies car images into six viewpoints (Front, Front-Left, Front-Right, Rear, Rear-Left, Rear-Right) with an additional "None" class for irrelevant or ambiguous images. It includes a robust pipeline for preprocessing, model training, and optimization for edge deployment.

---

Tasks and Milestones

1. Project Setup:

   • Create the GitHub repository.
   • Define the project structure and dependencies.

2. Data Preparation:

   • Analyze and clean the dataset.
   • Develop a data loader and augmentation pipeline.

3. Model Development:

   • Design the initial model and implement training and validation loops.
   • Experiment with initial architectures and changing hyperparameters.
   • Add tool for faster convergence

4. Iterative Refinement:

   • Address class imbalances and improve performance on edge cases.
   • Fine-tune augmentations and model configurations.

5. Model Optimization:

   • Convert into optimized format i.e. TFLite.
   • Test for latency and accuracy.

[x] Concluded

---

Setup

1. Clone

git clone https://github.com/omrastogi/car_part_detection.git
cd car_part_detection

2. Install Packages

conda create --name viewpointcq python=3.11 -y
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

3. Download Models

mkdir -p lite_models/viewpoint  # Ensure the directory exists
wget -P lite_models/viewpoint https://huggingface.co/omrastogi/viewpoint/resolve/main/convnext.tflite

---

Inference

python test_predict.py \ 
--folder data/5e9112c35026365e15eb871b \
--model lite_models/convnext.tflite \
--csv results.csv

Training

Download Prepared Dataset

Find the preprocessed dataset here: https://huggingface.co/datasets/omrastogi/car_parts_dataset

Prepare Dataset (Optional)

python scripts/preprocess_ann.py

Training

CUDA_VISIBLE_DEVICES=0 python scripts/train.py --base_data_dir data \
                --train_data_path data/training_data.json \
                --val_data_path data/val_data.json \
                --batch_size 8 \
                --learning_rate 0.0005 \
                --num_iterations 10000 \
                --log_interval 5 \
                --val_interval 50 \
                --checkpoint_interval 1000 \
                --grad_accumulation 4 \
                --unfreeze_interval 400 \
                --resize_to 380 380 \
                --device cuda

---

Training

Validation

Vaidation F1 Score

---

Evaluation

python scripts/evaluate.py --test_data_path data/test_data.json \
                   --checkpoint_path checkpoints/convnext/checkpoint_iter_5000.pth \
                   --base_dir data \
                   --resize_to 380 380 \
                   --batch_size 32

Overall Metrics:
Accuracy 0.86
F1 Score 0.85

Evaluation Report

---

TFLite Conversion

Convert a model to tflite in the following way

python src/tflite/convert_to_tflite.py \
    --checkpoint_path checkpoints/checkpoint_iter_5000.pth \
    --output_path lite_models/convnext.tflite \
    --input_size 380 \
    --num_classes 7 

Inference the tflite

python src/tflite/infer_tflite.py \
   --model_path lite_models/convnext.tflite \ 
   --image_path data/5f4dd0caf0b0b46649993480/scraped_0n61nM_1598934146856.jpg

---

Key Challenges

• Handling ambiguous and overlapping viewpoints.
• Ensuring robust performance for the "None" class with limited data.
• Handling the imbalance in the between the classes

---

Future Scope

• Extend the model for additional car viewpoints or object detection.
• Incorporate real-time feedback mechanisms for further deployment enhancements.

---

Extras

Create requirements.txt:

pip freeze | findstr /V " @ file://" > requirements.txt

Training Runs:

1. MobileNetV2: https://wandb.ai/omegam/my_car_classification_project/runs/x06uzfsc?nw=nwuseromegam

2. EfficientNetB4: https://wandb.ai/omegam/my_car_classification_project/runs/toev44gf?nw=nwuseromegam

---