ewrfcas / MST_inpainting

Learning a Sketch Tensor Space for Image Inpainting of Man-made Scenes (ICCV 2021)
MIT License
83 stars 11 forks source link

作者,您好,请问 LSM-HAWP wireframe detection model 仅仅是在ShanghaiTech数据集训练的而在所有实验数据集上公用的吗? #5

Open micklexqg opened 2 years ago

micklexqg commented 2 years ago

如题,如果LSM-HAWP wireframe detection model 只在ShanghaiTech数据集训练,不在其他数据集上训练,会不会影响在其他数据集上的实验效果?谢谢!

ewrfcas commented 2 years ago

我们只在ShanghaiTech上训练,在其他数据集上表现也比较良好。因为主要也是针对建筑方面的情况,自然场景直线比较少。另外可以借鉴SOLD2这种自监督方式来到其他数据集上finetune。 https://openaccess.thecvf.com/content/CVPR2021/papers/Pautrat_SOLD2_Self-Supervised_Occlusion-Aware_Line_Description_and_Detection_CVPR_2021_paper.pdf

micklexqg commented 2 years ago

谢谢!有个困惑,为啥论文里的在 comprehensive Places2 (P2C) (随机挑选的10个类别带自然场景和man-made场景的数据)比在man-made Places2 (P2M)(10个类别只包含man-made场景的数据)上效果好(Table 1),man-made Places2 (P2M)数据更贴近建筑方面,效果应该更好才对

ewrfcas commented 2 years ago

这可能是因为P2M的数据结构较多,相比自然场景的P2C本身难度就更高的原因