Anshu J.

Key Skills

Software

Roboflow

Top Subject Matter

Computer Vision / Disease Detection

Top Data Types

Image

Top Task Types

Object Detection

Freelancer Overview

I have hands-on experience building and preparing custom datasets for real-world computer vision tasks. For my AI-Based Basketball Hoop Detection project, I created and annotated a custom dataset from scratch using Roboflow, applying augmentation techniques such as flipping, scaling, and color adjustments to improve model robustness and generalization. This end-to-end dataset pipeline — from raw image collection to labeled, augmented training data — directly supported training a YOLO model that achieved reliable real-time detection accuracy. Beyond annotation, my broader project work reflects a strong understanding of how data quality and pipeline design impact model performance. In my Smart Crop Monitoring system, I logged and managed 500+ real-world detection events with structured metadata in SQLite, reinforcing the importance of clean, well-organized training and evaluation data. Across projects involving RT-DETR, YOLO, and custom CV toolkits, I have consistently worked at the intersection of data preparation, model training, and evaluation — giving me practical insight into how labeled data choices affect benchmarks like mAP, inference speed, and deployment readiness.

Labeling Experience

Basket-Ball Hoop

ImageBounding Box

Data Labeling Tasks Performed: Collected and curated a custom image dataset of basketball hoops across varied lighting conditions, angles, and court backgrounds. Used Roboflow's annotation tool to manually draw and label bounding boxes around the hoop region for each image, designating it as the primary detection class. Leveraged Roboflow's built-in augmentation engine to apply transformations including horizontal flipping, brightness/contrast adjustment, scaling, and rotation — expanding the dataset and improving model generalization across diverse real-world scenarios. Quality Measures Adhered To: Maintained strict bounding box consistency — ensuring tight, accurate annotations with minimal label noise throughout the dataset. Used Roboflow's dataset health check features to detect and resolve issues such as duplicate images, missing labels, and class imbalance before training. Validated final model performance through real-time inference testing to confirm detection reliability prior to deployment.

2025 - 2025

Knowledge Distillation—Object Detection Pipeline (Custom Dataset Benchmarking)

OtherImageObject Detection

Built an object detection dataset pipeline and trained YOLO/RT-DETR-style models using a custom dataset. Focused on optimizing model performance and inference efficiency by comparing detection accuracy and runtime metrics. Performed evaluation to validate improvements from model compression and knowledge distillation across architectures. • Used custom datasets for disease and general object detection experiments. • Benchmarked multiple CNN backbones (YOLO, EfficientNet-B0, ResNet-50) using mAP, inference speed, and model size. • Applied master–slave knowledge distillation to train RT-DETR-18 from RT-DETR-50. • Validated lightweight deployment readiness with evaluation results.

2021 - 2022

Education

Work History