Rico's Nerd Cluster

「离开世界之前一切都是过程」

[ML] FoundationPose (CVPR 2024)

Introduction Code, paper To reduce manual effort for large-scale training, FoundationPose introduces a synthetic data generation pipeline built on 3D model databases (G...

Posted by Rico's Nerd Cluster on May 26, 2026

[Robotics] From Market-Ready ROVs to Low-Cost AUVs: Building Practical Underwater Autonomy

This is a summary of From Market-Ready ROVs to Low-Cost AUVs BlueROV2 is a small 6-DoF ROV ROV platform that can operate at depths up to 100 meters, weighs under 12 kg, and includes a tether of up...

Posted by Rico's Nerd Cluster on May 8, 2026

DETR

Breakthroughs of RF-DETR (2026) 这不是小幅迭代，这是质的飞跃。 60 mAP，意味着它能识别的物体更准、漏检更少、误报率更低。而6.8ms的延迟，意味着它可以实时跑在普通GPU上。这两个指标同时满足，业界等了整整五年。你可能有个疑问：为什么这次突破来自DETR家族，而不是YOLO系列？因为架构路线不同。 YOLO走的是CNN路线，擅长速度...

Posted by Rico's Nerd Cluster on May 3, 2026

DETR

Introduction Here is the general architecture of DETR, which is quite straight-forward: First is a CNN to extract features (256-Vector) Second is a transformer to learn bounding boxes usin...

Posted by Rico's Nerd Cluster on May 1, 2026

Data Augmentation For Object Detection

TODO YOLOv8 employs a series of data augmentation techniques, including Mosaic, Mixup, random perspective transformation, and HSV augmentation. Additionally, the input images undergo meticulous pr...

Posted by Rico's Nerd Cluster on April 30, 2026

FEB-YOLOv8

FEB-YOLOv8 tries to make YOLOv8n small enough for underwater robots, then gives back the lost accuracy using attention and better small-object feature fusion. The three main changes are P-C2f, EMA ...

Posted by Rico's Nerd Cluster on April 28, 2026

Muon SGD

Muon means MomentUm Orthogonalized by Newton–Schulz. In practice, it takes the usual SGD momentum update for a 2D weight matrix, then orthogonalizes that update before applying it. NVIDIA’s Emergin...

Posted by Rico's Nerd Cluster on April 19, 2026

Progloss

ProgLoss is a training-time loss schedule for object detection. Instead of keeping classification and localization weights fixed for all epochs, it changes them over training progress: \[L_{\opera...

Posted by Rico's Nerd Cluster on April 16, 2026

Small Target Label Assignment (STAL)

STAL STAL = Small-Target-Aware Label Assignment. It is a training-time matching rule for object detectors: it decides which predicted boxes / anchors should be treated as positive samples for each...

Posted by Rico's Nerd Cluster on April 16, 2026

Target Label Assignment (STAL)

TOOD

TAL TAL usually means Task-Aligned Label Assignment or Task Alignment Learning in object detection. It is used during training, not inference. Its job is to decide: \[\text{Which anchors / predi...

Posted by Rico's Nerd Cluster on April 16, 2026