64种视觉骨干在树木分割与立体深度上的受控对比
原标题：What Does the Encoder Actually Decide? A Controlled Comparison of Vision Backbones on Joint Tree Segmentation and Stereo Depth
AI 摘要
该研究在固定数据集、解码器、损失函数和训练计划的前提下，仅替换编码器，对比了64种视觉骨干网络在树木语义分割与立体深度估计联合任务上的表现。结果显示，卷积和混合编码器（如internimage_t）表现最佳，而4个普通视觉Transformer中有3个从零训练时崩溃；参数量并不能预测质量，仅1.2M参数的edgenext_xxs优于大近两个数量级的模型；分割与深度排名高度一致，说明共享编码器不存在任务冲突。此外，64个编码器中有25个退化为全树分割，区域IoU无法发现但边界F1能暴露。
正文节选
What Does the Encoder Actually Decide? A Controlled Comparison of 64 Vision Backbones on Joint Tree Segmentation and Stereo Depth Abstract A robot that prunes trees needs two facts about the same pixel: whether it belongs to a tree, and how far away it is. Both are normally obtained by attaching a task head to a vision backbone, and the backbone is chosen by reputation rather than by measurement. We ask a narrow but answerable question: with the dataset, the decoders, the losses, the schedule an