Rico's Nerd Cluster

「离开世界之前一切都是过程」

[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

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

Neural Processing Unit

An NPU is a Neural Processing Unit. It is a specialized chip designed to run neural network inference efficiently, especially on edge devices like phones, drones, cameras, robots, and embedded syst...

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

Review of YOLO Models

Metrics mAP 50 - 95 For each IoU threshold from 0.50 to 0.95, we decide whether each predicted bbox is a true positive by checking whether its IoU with a matched ground-truth box is at least that...

Posted by Rico's Nerd Cluster on April 14, 2026

[BEV] DETR3D

DETR3D (CoRL 2021) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 6 camera images + calibration ↓ Shared 2D backbone ResNet / VoVNet ↓ FPN multi-scale 2D image feature...

Posted by Rico's Nerd Cluster on April 11, 2026