subhrokomol/siglip2-large-lora-v1-dataset
收藏资源简介:
该数据集名为SigLIP2-Large LoRA v1 — Training Pairs,是一个用于微调SigLIP2-large模型的材料表面视觉相似性训练对数据集。它通过LoRA(低秩适应)和监督对比损失进行训练优化。数据集中的每一行数据将一个目录产品图像与从SAM3分割的内部房间渲染中提取的多边形裁剪材料作物相关联。数据生成过程包括:对内部房间渲染进行SAM3分割,使用voyage-multimodal-3.5生成每个片段的多模态嵌入,执行近似最近邻搜索匹配产品图像目录,并保留通过相似性阈值的top-k匹配。数据集包含5,618个训练样本和576个评估样本,文件包括JSONL格式的数据文件和包含图像片段的压缩包。数据集支持视觉相似性任务和材料识别研究,适用于对比学习和监督对比学习场景。
The dataset is named SigLIP2-Large LoRA v1 — Training Pairs, which is a material surface visual similarity training pair dataset for fine-tuning the SigLIP2-large model. It is trained and optimized using LoRA (Low-Rank Adaptation) and supervised contrastive loss. Each row in the dataset associates a catalog product image with polygon-cropped material crops extracted from SAM3-segmented indoor room renders. The data generation workflow includes: conducting SAM3 segmentation on indoor room renders, generating multimodal embeddings for each segment via voyage-multimodal-3.5, performing approximate nearest neighbor search to match catalog product images, and retaining top-k matches that pass the similarity threshold. The dataset contains 5,618 training samples and 576 evaluation samples, with files including JSONL-formatted data files and a compressed package containing image segments. This dataset supports visual similarity task and material recognition research, and is applicable to contrastive learning and supervised contrastive learning scenarios.




