machine learning

3 articles

azure c++ computer vision cpp data analytics decorator deep learning game server geospatial data geosql GPU image segmentation Langchain LLM LM Studio machine learning multithreading network programming object oriented programming oop Optuna python pytorch RAG segformer synapse analytics tcp threads udp voice recognition

<Computer Vision> Implementing RTMDet-Ins 1. Backbone/Neck/Head Implementation

Sunday, Mar 30, 2025

Implementing Instance Segmentation Model From Basic Pytorch: Segformer I’m currently working on a pixel sprite detection pipeline called nino : For the segmentation model, I chose to implement RTMDet-Ins, because I haven’t used this model before, so it will be a good chance to learn about it It is the current SOTA model for instance segmentation in real time. 0. Overview paper link Need custom CSP backbone blocks PAFPN Heads for various tasks seems simple.

Jun Yeop(Johnny) Na

<Computer Vision> Implementing RTMDet-Ins 2. Making COCO Dataset Reader

Sunday, Mar 30, 2025

The most widely used image dataset format is COCO. COCO format: https://cocodataset.org/#home label metadata is handled in train/valid/test.json image metadata in images, annotation data in annotations, labels in categories 1. Basic COCO Format Example images: Contains image id and path where image is stored annotations: Contains image id and category id that the annotation belongs in. Can have multiple annotations(=instances) for each image labels include: category(label_id for classification) bbox([x_box_leftupper_corner, y_box_leftupper_corner, w, h], for object detection) and segmentation([x1 y1 x2 y2 … format]) categories: category id and the label’s name 2.

Jun Yeop(Johnny) Na

<Computer Vision> Implementing RTMDet-Ins 5. Pretraining Backbone on ImageNet

Sunday, Mar 30, 2025

Object Detection Model typically gets trained in two stages: Trains only backbone on classification task so that the backbone can provide reasonable output first. Trains the whole model on object detection task. This is because the detection head can’t train if the backbone isn’t providing useful embedding of the given image. The backbone’s output will give a random output to the head, which will lead the head to train in the wrong direction.

Jun Yeop(Johnny) Na