• This is our team mentor's github account

SIH2023-23

Team Terrainto

Problem Statement (SIH1418)

Description - Vision based methods using deep learning such as CNN to perform terrain recognition (sandy/rocky/grass/marshy) enhanced with implicit quantities information such as the roughness, slipperiness, an important aspect for high-level environment perception.

• Deployed ML model link: https://terrain.streamlit.app/

Python Libraries:

• Numpy
• Pandas
• Matplotlib
• Albumantations
• Torch
• Torchvision
• PIL
• Pickle
• Streamlite

Machine Learning Model:

• ResNet18 (pre-trained) with one Fully Connected Layer(512 x 4) predicting 4 classes

Dataset

• 31517 training images with 4 classes namely Grassy, Marshy, Rocky, Sandy
• 6765 validation images and 6769 test images
• Normalized with the pretrained ResNet18 measures.

Training

• The pre-trained layers are freezed and only the last FC layer was trainined for 1 complete pass through all the samples.
• Mini_batch size: 32

Validation

• Accuracy on validation_set 91.1 % and on test_set 93.8%
• Mini_batch size: 32

Learning Curve

Integrating Web-Development

• The model is deployed using Streamlit library
• we Pickeled the model for its interation with Streamlit's "app.py"
• Image uploader coloumn that uploads image to model and recieves predicted index as return value
• The result is displayed on the website.

Contributors

• Manav Jain (Leader): https://github.com/manavjain2005
• Harsh Singh : https://github.com/hharshas
• Abhinav Jain : https://github.com/jainjabhi05
• Aadhya Jain : https://github.com/aadhya0002
• Dyuti Ballav Paul : https://github.com/dyuti01
• Pratham Todi : https://github.com/pra1ham28