Conference- AAAI'25 - Stable Mean Teacher for Semi-supervised Video Action Detection- AAAI'24 - Semi-supervised Active Learning for Video Action Detection- CVPR'22 - End-to-End Semi-Supervised Learning for Video Action DetectionWorkshop- NeurIPSW'23 - Benchmarking self-supervised video representation learning- CVPRW'22 - Video Action Detection: Analysing Limitations and ChallengesReviewer - ICLR'23 '24 '25, CVPR'23, '24 '25, ECCV/ICCV'22 '23 '24, NeurIPS '23, '24

- Problem Statement: Open-Vocabulary Video Object Detection- Solution: Global-Local Soft Prompting. Devised approach via LLMs to associate visual and motion cues respective to different objects to help recognize objects for in-the-wild settings.

- Fine grained classification of bird species.- Segemntation and bird parts detection for improved classification.- N-shot learning

- Detecting Deepfakes with Metric Learning- 8th International Workshop on Biometrics & Forensics (IWBF'20)

Syn2Real: Forgery Classification via Unsupervised Domain Adaptation (IEEE WACVW'20) • Detection of Copy-Move Forgery Detection on CASIA and CoMoFoD datasets.• Semantic painting and Copy-Move Forgery techniques on the COCO dataset to generate a synthetic dataset of 80,000 images.• Used Unsupervised Domain Adaptation to improve the classification accuracy by adapting the feature space of realistic domain from a synthetic domain. This approach can be generalized to any field.

- Objective: Object detection model to recognize & improve the road safety of autonomous vehicles- RPN training to improve the feature extraction from images- Implemented Faster R-CNN, Mask R-CNN, RetinaNet & EfficientDet

- Hierarchical Classification of Indian Monuments.

- 11th Indian Conference on Computer Vision, Graphics and Image Processing Workshops (ICVGIPW'18). - Evaluated Graph-based Visual Saliency, Deep Local Features models for Visual Similarity between Indian monuments.- Improved the accuracy of the model by 4-5% using ensemble learning of supervised feature classification, salient detections and Inception trained networks.

Machine Learning and Python coding problems specifically in Computer Science Subject

- ICCVIP'18 CHALLENGE WINNER- 11th Indian Conference on Computer Vision, Graphics and Image Processing Workshops (ICVGIPW'18). - Interspecies classification with only 150 DSLR images. Analyzed differentdata augmentation techniques to countereffect the problem of an imbalanced dataset.- Analyzed the task with Transfer Learning and multi-stage training.- Achieved F1-score of 55.67% developing an end-to-end model of Mask R-CNN and ImageNet models and incorporating ensemble learning.

• Content-Based Video Relevance Feedback system.• Developed a Video Recommender system using frame-level features (InceptionV3 network) and video-level features (Convolutional 3D model) to solve the problems of ”cold-start” videos.• Evaluated Siamese Network, Deep Learning based Regression and Triplet loss.

• Developed an online system using NVIDIA Jetson TX1 to track pedestrians on road.• Trained the NVIDIA Caffe DetectNet and Pednet Model on DIGITS server to detect and track animals using COCO dataset.

• Founding Team Member at Start-up• Extracted the dataset of houses from Google Maps to analyze data.• Identified individual rooftops using Google Maps satellite imagery.• Achieved placement of Solar panels on optimal location in rooftop area.Orientation angle required was calculated using sun movement.

• Automated update of fuel level of the go-kart with sensor alert for the low levelof fuel was created using Ultrasonic Sensor and GSM Module.• According to the distance between obstacle and the kart, speed controlmechanism was devised.• Kill switch is controlled by interfacing between GSM Module, Enginetemperature & Fuel leakage system.• RPM data was monitored to check required speed for brake test.