Improving CLIP Training with Language Rewrites (2023)The researchers experimented with LLM-based text augmentation to improve CLIP model training, using 'Meta-Input-Output Text Pairs' as prompts for the LLM. The researchers generated four additional1 min read·nickonedrop.hashnode.devNov 6
Rethinking Lightweight Salient Object Detection via Network Depth-Width Tradeoff (2022)This paper splits the decoder into separate paths, each with a different job. The idea is to reduce semantic dilution, loss of spatial structure, and blurry boundaries in a typical U-shaped network.2 min read·nickonedrop.hashnode.devNov 5
S3OD: Towards Generalizable Salient Object Detection with Synthetic Data (2025)The idea is to build a synthetic training set for salient object detection (SOD), then train a model using a DINOv3 encoder with a custom decoder. Here’s how the reassembly step works, assuming a 224 min read·nickonedrop.hashnode.devNov 2
U²-Net: Going Deeper with Nested U-Structure for Salient Object Detection (2020)The idea behind the U-shaped structure is to keep local detail while also capturing the broader context needed for segmentation. What stands out about the proposed RSU-L block is that it works at mu1 min read·nickonedrop.hashnode.devOct 29
If AI Is Electricity, What Is the Refrigerator?AI is often compared to electricity. Electricity powered all kinds of new products, from factory machines to refrigerators and washing machines. The idea is that AI—foundation models like Claude and G2 min read·nickonedrop.hashnode.devOct 11