Jullian Yapeter

Conducted computer vision, deep learning, and controls R&D for this attraction. So happy to see it bringing joy to so many people!

Diverse multimodal dataset of 1085 teleoperated robot episodes. Includes multi-view camera observations, cartesian position and velocity data, and natural language annotations. Tasks include picking and placing of various items (of various categories) and receptacles.

Novel method to improve the performance of downstream unimodal language understanding tasks by leveraging multimodal data during training; employing optimization towards the task-specific objective while regularized against a multimodal teacher network in latent space.

A simple solution for multimodal representation learning, which takes advantage of modality-specific pre-trained encoders and a flexible, lightweight, latent-mixing network to effectively generate meaningful multimodal representations.

Application of a deep-learning model, Mask R-CNN, trained to identify healthy and Plasmodium-infected red blood cells. Capable of generating segmentation masks on top of bounding box classifications for immediate visualization and stage-specific identification. Potential to reduce errors common in manual counting through subsequent standardization.

Pre-trained predictive models on various tasks to facilitate RL training on a novel downstream task of the same distribution. The resulting reward-induced representation proved useful in both speeding up downstream PPO training, and improving final rollout performance.

AI-powered water monitoring system. Measures parameters of its surrounding environment, collects samples of the pollutants, and performs data analytics in the cloud to produce valuable information.

Final project for 30.007 Engineering Design Innovation course at SUTD. An ultrasonic-based wearable for the blind, using 3D positional audio feedback to assist in obstacle avoidance, and IMU-based fall detection and notification functionality over Twilio API. Won Best Demonstration Award at public project exhibition; awarded by industry guests.

Given a set of images with the same watermark, our method segments watermark boundaries (employing graph cut) and subsequently estimates the watermark's parameters (color + alpha-blending constant). It then removes the watermark using an inverse watermarking transform.

This one started because I wanted to colorize my grandparents' old portraits. Our AutoColorizer fine-tunes a VGG16 network pretrained on ImageNet. It is then trained on a dataset of 200,000 celebrity portraits. We formulate the colorization process as the prediction of the AB channels given the L channel of an image in the LAB color space.