Projects

• Proposed DroneSegNet, a semi-supervised network for robust segmentation by jointly utilizing aerial scenes and elevation maps.

• Introduced a spatial dependency module to augment features with elevation maps using bi-LSTMs.

• Achieved an F1 mean of 0.8048, ranking 4th on the DroneDeploy Aerial Segmentation Benchmark.

• Implemented classification and clustering algorithms (KNN, K-means and bisecting K-means) on the Shuttle dataset in Hadoop and Apache Spark.

The project was completed for partial fulfilment of the course Foundations of Data Science in BITS Pilani.

• Trained Mask R-CNN and Hybrid Task Cascade models for Instance Segmentation on the Indian Driving Dataset. Used mixed-precision and multi-scale training for better feature understanding.

• Trained DeepLabv3+ on the Indian Driving Dataset for Semantic Segmentation. Used Gradient accumulation for optimizing batch size vs crop size trade-off.

• Achieved 16th place finish in ICCV's AutoNue 2019 challenge.

• Implemented parallel bi-LSTM and bi-GRU models on the Liar-Plus dataset for binary and multi-class classification of news captions.

• Achieved 59% binary classification and 25% sixway classification accuracy.

• Implemented five Deep Learning architectures (LeNet-5, AlexNet, VGG16, VGG19 and ResNet18) to classify German Traffic Signs (GTSRB dataset) using Keras.

• Achieved 95% test accuracy using a modified LeNet architecture.

• Provided a working solution to gauge price ranges for HackerEarth's IndiaMART Hackathon.

• Used Z-score based outlier detection for an accurate "average estimation" on each unit of measurement.

• Trained classification and object detection models for Early Blight detection in tomato leaves using Rajasthan sample images.

• Completed Proof-of-concept with 99.14% classification accuracy and 60% detection test recall.

• Developed a security solution for residential complexes to detect license plates at the gate and classify as residents and guests.

• Added features to invite guests (allowing them a seamless visit), school bus tracking and vehicle logs.

• Developed a Smart Lock to 'lock' the vehicle within the complex and raised suspicion alerts based on outlier detection to prevent unauthorized access or theft attempts.

• Came second at the finals of Smart India Hackathon 2019.

• Implemented a pipeline using OpenCV for detecting lane lines in images and videos of roads in Python for navigation in an autonomous vehicle.

• Used Canny edge detection to detect edges, followed by Hough transform to detect prominent lines in the Region of Interest.

• Used transformation to LAB colourspace to effectively handle low-light images.