Octav Florescu

• Improved traffic signs facing/orientation via Left-Right Classification and by using the angle estimated from the aspect ratio of the sign
• Being the key contact for partnership with BabesBolyaiUniv on signs-facing (financed) research for MSc students, I have prepared the data for their study and gave my input on research directions, supported and coordinated the researches in a joint effort with a coordinator from BBU
• Improved detector's dataset splitting method via hyperopt tries
• Extracted detections proximity relations Graph from object's locations in images, in order to further use them in predicting logical-groups and correct their 3D positions
• Extracted detector's roi embeddings from various layers of the fpn in a mmdetection cascadercnn, and used them as embedded info for another classification meta info extraction task
• Made the code dynamic such that any additional custom head can be inserted
• Productized Meta's SAM model for traffic signs segmentation
• That is dockerisation and adaptation to internal data infrastructure
• Trained a traffic signs detector via super-gradients and integrated instance segmentation train logic
• Integrated the resulted model in Kotlin for Android/IoT processing via ONNX, TfLiteGPU and Qualcomm's SNPE to run it on DSP
• Internal framework for multi-task/-head model training and quantized export
• Added obj detection, instance segmentation, various attributes classifications
• Autotagging pipeline plus visual prompting of GPT4o and Qwen2VL 72b for Traffic signs attributes and curating inference of previous components

2024

• Quectel GPS module continuous read write & access via flask API

2022

• trees detection and counting from airplane images in order to prevent deforestation https://github.com/ai-in-actiune/tree-counting-and-classification-in-images

2021

• Automatic triggered watering of plants system https://github.com/octavflorescu/SelenoidSoilWatering

2019

• Trained (TPU) and physical-assemble concept (3d printed the skeleton) of a GoogleCoral board for counting the pull-ups of each team member. https://github.com/octavflorescu/Coral-camera-pullups-counter

2016

• Whist Score Holder in swift - the UI/UX and implementation - https://github.com/octavflorescu/WhistScoreHolder

• Managed the processing pipeline between ML services running in Docker containers
• TrafficSigns' Positioning improvements by interpolating the images delayed provided locations and timestamps of the captured images and captured gps loc
• Other positioning improvements for rotated signs by using the difference between the sign's width and the physical sign's width, given the expected height
• Made a web tool for creating a geo-clustering dataset where signs detected in several images are clustered manually along with their actual north-heading (relevant for traffic signs positioning)
• Work on a TrafficSigns detection component using YoloV3/V4 darknet
• Custom dataset splitter based on region (gps) and minimum counts per each set
• Learning the best split random seeds via hyperopt
• Work on a TrafficSigns aggregator component: finding the best parameters for augmenting the tabular data pre-clustering using hyperopt
• Trained several RESNET based models for the task of multi class classification on traffic signs of Indonesia
• The main purpose was of upgrading a traffic sign-not traffic sign binary classification component to a multi class (not traffic sign & several TrafficSign types) component from the protobuf RabbitMQ pipeline
• Trained a custom pytorch cnn model and visualisations using gradcam on CIFAR10
• Technologies: Python, Pytorch, FastAI, RabbitMQ, Docker, Jenkins, Protobuf
• Pretrained models I experimented with: Resnet34/50/101, WideResnet50/101, AntialiasedResnet34/50, Inception v3, SeResneXt50
• Components I experimented with: SpatialTransformNetwork, SpatialPyramidPool

• Development on Scout GPS Link By Telenav Inc
• - US/Canada App store - iPhone full Navigation app and broadcasting the iOS screens to the Toyota/Lexus onboard screen
• SkinningKit - Internal UI-SDK for easier styling of UI elements
• It provided one entry point to recolor/reformat the whole app by applying another 'Skin' (colors/forms/shadows/roundedCorners)
• Experimented with UI components
• SGL2 - Not released
• ThinkNear's Background Location iOS SDK - Not released - It was supposed to track the user and make a profile of them for advertising so I researched a lot @GPS silent tracking
• Hybrid / GPS Navigation (Sat Nav) By Telenav GmbH
• Technologies: iOS, Objective-C, Swift
• Experimented with architectures, multithreading, route ETA estimation, CoreData, GPS, AppDelegate, CustomMapSDKs usage, UserDefaults, Requests, UI constraints, Xibs, Portrait SafeArea auto adaptation, Transition Animations

• Company Overview: https://www.facebook.com/MatrixRobots/
• Mega-Sumo Robots programmer
• Technologies: C, Arduino
• Https://www.facebook.com/MatrixRobots/