Md Awsafur Rahman

Top Scorer at IEEE Video & Image Processing Cup 2022 at ICIP'22 IEEE Nov 2022 The goal of this competition was to detect synthetic/fake images which are generated by Generative Models. The most challenging part of this competition was to generalize for "Unknown Generators". Our team came up with a novel solution that led us to 1st on the leaderboard. Our team outperformed other teams by a huge margin (+15% difference). We also compiled a large-scale dataset for detecting "Real Vs Fake" images which we wish to release along with our novel solution to the community. Sadly,… Show more The goal of this competition was to detect synthetic/fake images which are generated by Generative Models. The most challenging part of this competition was to generalize for "Unknown Generators". Our team came up with a novel solution that led us to 1st on the leaderboard. Our team outperformed other teams by a huge margin (+15% difference). We also compiled a large-scale dataset for detecting "Real Vs Fake" images which we wish to release along with our novel solution to the community. Sadly, our team was ranked 2nd by the judge for unknown reasons. Show less

Winner at 2022 Kaggle ML Research Spotlight: August Kaggle Aug 2022 Kaggle ML Research Spotlight award is intended to support ML researchers working to reproduce code as well as provide a valuable cutting-edge resource to the broader Kaggle community.In my work, I re-implemented the recently published paper "GCViT: Global Context Vision Transformer" by NVIDIA. Also, I explained the paper with annotated figures. Moreover, I created a live demo in Hugging Face using Gradio. Furthermore, I've created an open-source library with ported ImageNet pretrain… Show more Kaggle ML Research Spotlight award is intended to support ML researchers working to reproduce code as well as provide a valuable cutting-edge resource to the broader Kaggle community.In my work, I re-implemented the recently published paper "GCViT: Global Context Vision Transformer" by NVIDIA. Also, I explained the paper with annotated figures. Moreover, I created a live demo in Hugging Face using Gradio. Furthermore, I've created an open-source library with ported ImageNet pretrain weights so that it can be reused for other projects as well. Show less

Winner at 2022 Google OSS Expert Prize: May Google & Kaggle May 2022 The goal of this challenge was to publish open-source high-quality community content (either notebooks or discussions) using Tensorflow, TF Lite, Keras, JAX, or Flax. The contents were evaluated on three key metrics: Technical Expertise, Presentation, and Documentation. The proposed solution was on 2D projections of 3D image stacks as training data for segmentation using a TransUNet model. It showed how to use tf.keras, tf.data & tfrecords along with tf,keras for medical image segmentation.

Winner at IEEE Signal Processing Cup 2022 at IEEE ICASSP'22 IEEE Signal Processing Society May 2022 The goal of this competition was to identify the algorithm responsible for the deepfake speech. One of the key challenges was to identify Unknown algorithms which contain numerous unseen and out-of-distribution samples. Our solution includes tackling unseen and out-of-distribution data while maintaining a good score for known data distribution. The proposed solution was able to outrank 2nd and 3rd teams with over 10% accuracy difference.

Winner at 2022 Google OSS Expert Prize: January Google & Kaggle Jan 2022 The goal of this challenge was to publish open-source high-quality community content (either notebooks or discussions) using Tensorflow, TF Lite, Keras, JAX, or Flax. The contents were evaluated on three key metrics: Technical Expertise, Presentation, and Documentation. The proposed solution was on finding extraterrestrial signals in data from deep space (SETI Dataset) using TensorFlow. It showed how to use tf.keras, tf.data & tfrecords along with tf,keras to identify signals in deep space.

Winner at Deep Chimpact DrivenData Nov 2021 The goal of this challenge is to use Deep Learning to automatically estimate the distance between a camera trap and an animal in a series of camera trap videos. We had to predict the distance between the animal and the camera at each point in time. One of the main challenges of this competition was to tackle high deviation between Train and Test Distirubiton. Our solution outperformed the other teams and got a 1.6203 MAE score to win 1st Place.

Winner at KaggleDays x ZbyHP Championship Meetup in Shanghai Kaggle Nov 2021 The Championship consists of 12 online meetups with an in-person final event in Barcelona, Spain. The 3 winning teams' online meetup is invited to participate in person in the final and get a chance to become the ultimate Kaggle Days champion. Unlike any kaggle competitions, these competitions didn't last for 2-3 months rather 4hours which creates a very competitive environment. We had to make some critical decisions starting from modeling, data augmentation, loss function, etc within just… Show more The Championship consists of 12 online meetups with an in-person final event in Barcelona, Spain. The 3 winning teams' online meetup is invited to participate in person in the final and get a chance to become the ultimate Kaggle Days champion. Unlike any kaggle competitions, these competitions didn't last for 2-3 months rather 4hours which creates a very competitive environment. We had to make some critical decisions starting from modeling, data augmentation, loss function, etc within just 4hours. Our team got a 0.96039 F1score and become the winner jointly with another team having the same score. Show less

Best Student Team & 4th Place winner at SIIM-FISABIO-RSNA COVID-19 Detection Kaggle Aug 2021 In this competition, we had to identify and localize COVID-19 abnormalities on chest radiographs. In particular, we had to categorize the radiographs as negative for pneumonia or typical, indeterminate, or atypical for COVID-19. Our model worked with imaging data and annotations from a group of radiologists. Hopefully, it'll help radiologists diagnose the millions of COVID-19 patients more confidently and quickly. This will also enable doctors to see the extent of the disease and help them make… Show more In this competition, we had to identify and localize COVID-19 abnormalities on chest radiographs. In particular, we had to categorize the radiographs as negative for pneumonia or typical, indeterminate, or atypical for COVID-19. Our model worked with imaging data and annotations from a group of radiologists. Hopefully, it'll help radiologists diagnose the millions of COVID-19 patients more confidently and quickly. This will also enable doctors to see the extent of the disease and help them make decisions regarding treatment. In this competition, we won the Best Student Team Award and also got the first Gold Medal in a Kaggle Competition achieving 4th place globally. Show less

Winner at DhakaAI: Traffic Detection Green University of Bangladesh Dec 2020 DhakaAI-2020 brought the university and college students from the local and international arena to participate in an open AI-based object detection challenge and get their first experience through solving a real-life problem in a competitive manner.In this competition, we had to tackle some major issues, starting with highly dense vehicles, day-night imbalance, class imbalance, etc. Our solution overcome these issues and outperformed other teams and won the 1st place.

First Runner-Up at IEEE Video & Image Processing Cup at ICIP'2020 IEEE Signal Processing Society Oct 2020 The 2020 VIP-cup challenge focused on fisheye cameras mounted into street lamps at junctions and vehicle detection and tracking to be used for a junction management system to optimize the flow of traffic and synchronize with other junctions to obtain bottleneck performances throughout the city. Fisheye cameras are used since they tend to be promising in terms of reliability and scene coverage at a chosen junction. They provide 360 degrees of observation view, thus introducing key changes in… Show more The 2020 VIP-cup challenge focused on fisheye cameras mounted into street lamps at junctions and vehicle detection and tracking to be used for a junction management system to optimize the flow of traffic and synchronize with other junctions to obtain bottleneck performances throughout the city. Fisheye cameras are used since they tend to be promising in terms of reliability and scene coverage at a chosen junction. They provide 360 degrees of observation view, thus introducing key changes in traffic management.This competition challenged us with critical issues such as High distortion ratios, Different scales of same target object moving in different parts of the image, Day/night views variance (night view suffers from low quality related to surrounding lightning conditions), Exposure introduced with vehicle lights (night view). We had to overcome all of these issues to achieve the 2nd position. Show less