Team Cuda – Annam AI Hackathon Preliminary Round Submission

---

Video Link: Google Drive Folder

---

Team Members

• K V Srinanda
• M Manvith Prabhu
• N Dhruva
• Abhayjit Singh Gulati
• Aryan Herur

---

Task 1: Soil Image Classification

Approach

• Used PyTorch with a pre-trained DenseNet for transfer learning.
• Applied data augmentations using torchvision.transforms.
• Split data into 80% training and 20% validation.
• Fine-tuned the model and retrained on the full dataset for better generalization.

Challenges

• Faced class imbalance, with some soil types underrepresented.

Solution

• Used weighted cross-entropy loss to assign higher penalties to minority class errors, improving prediction balance.

Results

• Achieved an F1 Score of 1.0 on the public leaderboard.

---

Task 2: Anomaly Detection (Soil vs Non-Soil)

Approach

• Extracted semantic features using DINOv2 ViT-S/14.
• Trained a One-Class SVM using only soil images.
• Standardized features with StandardScaler.

Challenges

• One-class training data made it difficult to model the decision boundary.

Solution

• Used robust DINOv2 embeddings and tuned SVM hyperparameters (nu, gamma) to avoid overfitting.
• Applied feature normalization to improve model stability.

Results

• Achieved a public leaderboard score of 0.9326.

---

Setup and Run Instructions

1. Clone the repository

git clone https://github.com/abhayjit07/teamcuda_annam.git

2. Create and activate a virtual environment

python3 -m venv env
source env/bin/activate  # On Windows: env\Scripts\activate

3. Install dependencies

pip install -r requirements.txt

4. Download and prepare the dataset

bash data/download.sh

5. Open the notebooks training_task1.ipynb/training_task2.ipynb and run all cells to train models for respective tasks.

6. Open the notebooks prediction_task1.ipynb/prediction_task2.ipynb and run all cells for evaluation.