IIITH Campus Image Prediction

This repository contains a collection of models designed to localize images taken at various locations across the IIIT-Hyderabad campus. The pipeline includes region classification, viewing angle prediction, and geographic coordinate estimation.

📌 Overview

The system comprises three major components:

1. Region Classification
   Classifies an image into one of 15 predefined campus regions using a fine-tuned CNN.

2. Viewing Angle Prediction
   Predicts the direction the camera was facing (angle from North) using region-specific regressors.

3. Coordinate Prediction
   Estimates the (latitude, longitude) location of the image by classifying it to a representative anchor within the predicted region.

---

📂 Notebooks

1. Region Classification

Notebook: region_classification.ipynb
Trains a ConvNeXtV2 Tiny model (convnextv2_tiny.fcmae_ft_in1k) for 15-way classification of images into campus regions.

• Training Strategy: Fine-tuned on 6000+ campus images using CrossEntropyLoss.
• Preprocessing: Resize to 224×224, normalize with ImageNet stats, and apply augmentations (flips, crops).
• Output: Region ID (1–15)

2. Viewing Angle Prediction

Notebook: angle_prediction.ipynb
Implements a two-stage ensemble for predicting the viewing direction in degrees:

• Stage 1: A frozen ConvNeXtV2 model classifies the image region.
• Stage 2: A region-specific ConvNeXtV2 model regresses the angle in [cos, sin] form using MSELoss.
• Postprocessing: Angle recovered using atan2(sin, cos).

3. Coordinate Prediction (Anchor Classification)

Notebook: coordinate_prediction.ipynb
Uses fused features and anchor classification for coordinate prediction:

• Feature Extraction: Extracts and fuses features from frozen ConvNeXtV2 (CNN) and ViT (Transformer).
• Anchor Setup: Applies K-Means clustering to region-wise coordinates to define N anchor points per region.
• Classification: Trains a region-specific MLP using CrossEntropyLoss to classify the fused feature into one of the region's anchor indices.

---

🧠 Key Ideas

• Trained region-specific expert models for angle and coordinate prediction, guided by a shared region classifier.
• Converted coordinate regression to a classification task via anchor clustering.
• Combined features from CNNs and Transformers to improve semantic understanding and localization accuracy.