CFM中2D到3D相似度低解决方案

type

status

date

slug

summary

📝 主旨内容

改变输入

💡

多视角输入，通过3D点云生成多视角图片

改变架构

💡

卷积，KAN，KAN卷积，类重构先降维再升维（降维损失+重构损失）

结果

论文结果(投影)	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	0.990	0.894	0.986	0.989	0.980	0.916	0.951	0.916	0.986	0.886	0.949
AUPRO@30%	0.979	0.963	0.982	0.940	0.944	0.961	0.980	0.983	0.972	0.980	0.968
AUPRO@10%	0.937	0.892	0.947	0.890	0.838	0.885	0.940	0.948	0.918	0.941	0.914
AUPRO@5%	0.878	0.806	0.894	0.830	0.742	0.799	0.882	0.897	0.853	0.882	0.846
AUPRO@1%	0.469	0.402	0.486	0.450	0.380	0.397	0.463	0.490	0.453	0.463	0.445

复现投影	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	0.995	0.865	0.980	0.993	0.937	0.868	0.967	0.934	0.947	0.864	0.935
AUPRO@30%	0.980	0.959	0.982	0.943	0.922	0.962	0.980	0.983	0.969	0.981	0.966
AUPRO@10%	0.939	0.881	0.947	0.897	0.774	0.891	0.941	0.948	0.908	0.944	0.907
AUPRO@5%	0.881	0.791	0.895	0.841	0.671	0.808	0.884	0.896	0.839	0.888	0.839
AUPRO@1%	0.469	0.389	0.489	0.487	0.342	0.406	0.465	0.490	0.447	0.475	0.446
P-AUROC	0.997	0.988	0.999	0.972	0.977	0.992	0.998	0.998	0.997	0.996	0.991
cos_sim_2Dto3D	0.990	0.987	0.988	0.986	0.988	0.983	0.989	0.990	0.984	0.983	0.987
cos_sim_3Dto2D	0.820	0.876	0.833	0.791	0.880	0.821	0.808	0.839	0.849	0.854	0.837

复现投影(author checkpoint-50ep)	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	0.994	0.874	0.943	0.987	0.817	0.787	0.940	0.877	0.625	0.702	0.855
AUPRO@30%	0.980	0.952	0.982	0.945	0.929	0.932	0.980	0.982	0.952	0.967	0.960
AUPRO@10%	0.940	0.861	0.946	0.900	0.797	0.817	0.941	0.946	0.859	0.903	0.891
AUPRO@5%	0.882	0.766	0.893	0.846	0.695	0.714	0.882	0.892	0.758	0.817	0.815
AUPRO@1%	0.469	0.369	0.483	0.498	0.343	0.349	0.463	0.486	0.358	0.394	0.421
P-AUROC	0.997	0.985	0.998	0.972	0.979	0.984	0.998	0.998	0.993	0.990	0.989
cos_sim_2Dto3D	0.989	0.986	0.986	0.985	0.980	0.979	0.988	0.988	0.968	0.976	0.983
cos_sim_3Dto2D	0.799	0.856	0.803	0.759	0.786	0.772	0.776	0.795	0.617	0.757	0.772

复现投影更换数据集(author checkpoint-50ep)	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	0.990	0.903	0.980	0.990	0.973	0.898	0.948	0.916	0.972	0.858	0.943
AUPRO@30%	0.979	0.962	0.982	0.942	0.944	0.961	0.980	0.983	0.973	0.980	0.969
AUPRO@10%	0.939	0.890	0.947	0.895	0.837	0.885	0.941	0.949	0.921	0.940	0.914
AUPRO@5%	0.879	0.803	0.894	0.840	0.740	0.798	0.883	0.897	0.857	0.880	0.847
AUPRO@1%	0.466	0.403	0.486	0.488	0.376	0.396	0.466	0.494	0.455	0.469	0.450
P-AUROC	0.997	0.989	0.999	0.972	0.985	0.992	0.998	0.998	0.998	0.996	0.992
cos_sim_2Dto3D	0.990	0.987	0.988	0.986	0.988	0.984	0.989	0.990	0.984	0.983	0.987
cos_sim_3Dto2D	0.823	0.880	0.836	0.802	0.889	0.813	0.815	0.836	0.879	0.870	0.844

复现投影更换数据集(author checkpoint-250ep)	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	0.992	0.915	0.980	0.990	0.986	0.897	0.959	0.946	0.986	0.930	0.958
AUPRO@30%	0.979	0.965	0.982	0.941	0.952	0.964	0.980	0.983	0.975	0.982	0.970
AUPRO@10%	0.938	0.899	0.947	0.894	0.859	0.895	0.941	0.949	0.926	0.946	0.919
AUPRO@5%	0.879	0.818	0.893	0.838	0.768	0.813	0.883	0.897	0.865	0.892	0.855
AUPRO@1%	0.462	0.418	0.483	0.488	0.392	0.406	0.463	0.491	0.462	0.481	0.455
P-AUROC	0.997	0.990	0.999	0.971	0.987	0.993	0.998	0.999	0.998	0.997	0.993
cos_sim_2Dto3D	0.990	0.988	0.988	0.986	0.989	0.984	0.989	0.990	0.983	0.984	0.987
cos_sim_3Dto2D	0.823	0.891	0.825	0.799	0.892	0.816	0.814	0.822	0.882	0.873	0.844

0.0.

论文结果(卷积)	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	0.997	0.866	0.990	0.993	0.989	0.927	0.979	0.897	0.990	0.918	0.955
AUPRO@30%	0.979	0.965	0.982	0.941	0.948	0.969	0.982	0.983	0.977	0.981	0.971
AUPRO@10%	0.938	0.897	0.947	0.893	0.847	0.906	0.945	0.948	0.931	0.944	0.920
AUPRO@5%	0.880	0.813	0.894	0.834	0.756	0.820	0.891	0.896	0.872	0.889	0.855
AUPRO@1%	0.469	0.409	0.488	0.453	0.393	0.409	0.477	0.488	0.467	0.473	0.453

仅替换3Dto2D为卷积	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	0.992	0.877	0.981	0.996	0.996	0.938	0.982	0.923	0.954	0.881	0.952
AUPRO@30%	0.980	0.965	0.982	0.944	0.944	0.923	0.981	0.983	0.973	0.980	0.966
AUPRO@10%	0.941	0.897	0.947	0.898	0.898	0.777	0.943	0.948	0.919	0.941	0.911
AUPRO@5%	0.884	0.815	0.894	0.842	0.842	0.674	0.887	0.896	0.854	0.882	0.847
AUPRO@1%	0.473	0.410	0.486	0.477	0.477	0.344	0.468	0.491	0.458	0.469	0.455
P-AUROC	0.997	0.990	0.999	0.973	0.973	0.978	0.998	0.998	0.998	0.996	0.990
cos_sim_2Dto3D	0.990	0.987	0.988	0.986	0.986	0.988	0.989	0.990	0.984	0.983	0.987
cos_sim_3Dto2D	0.826	0.880	0.842	0.805	0.805	0.885	0.814	0.845	0.883	0.859	0.844

全替换为卷积(复现)	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	0.994	0.880	0.980	0.996	0.940	0.893	0.992	0.912	0.971	0.914	0.947
AUPRO@30%	0.979	0.963	0.982	0.945	0.923	0.968	0.981	0.983	0.975	0.982	0.968
AUPRO@10%	0.939	0.894	0.946	0.900	0.779	0.906	0.945	0.948	0.927	0.946	0.913
AUPRO@5%	0.881	0.809	0.893	0.843	0.678	0.826	0.890	0.897	0.864	0.893	0.847
AUPRO@1%	0.470	0.406	0.485	0.474	0.348	0.412	0.476	0.491	0.464	0.487	0.451
P-AUROC	0.997	0.989	0.998	0.973	0.977	0.992	0.998	0.998	0.998	0.996	0.992
cos_sim_2Dto3D	0.990	0.987	0.988	0.986	0.990	0.988	0.989	0.990	0.988	0.987	0.988
cos_sim_3Dto2D	0.825	0.881	0.845	0.805	0.884	0.827	0.817	0.843	0.884	0.858	0.847

全替换为encoder-decoder(复现)？？	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	0.997	0.841	0.983	0.989	0.938	0.890	0.966	0.891	0.950	0.915	0.936
AUPRO@30%	0.980	0.955	0.982	0.942	0.923	0.956	0.980	0.983	0.970	0.981	0.965
AUPRO@10%	0.940	0.871	0.947	0.895	0.777	0.879	0.940	0.948	0.911	0.944	0.905
AUPRO@5%	0.882	0.775	0.894	0.837	0.674	0.793	0.881	0.896	0.842	0.888	0.836
AUPRO@1%	0.473	0.382	0.487	0.471	0.343	0.399	0.461	0.489	0.447	0.475	0.443
P-AUROC	0.997	0.987	0.999	0.972	0.978	0.991	0.998	0.998	0.997	0.996	0.991
cos_sim_2Dto3D	0.990	0.986	0.988	0.986	0.988	0.983	0.989	0.990	0.984	0.982	0.987
cos_sim_3Dto2D	0.819	0.872	0.830	0.790	0.879	0.820	0.808	0.839	0.847	0.854	0.836

3D到2D替换为big-MLP	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	0.990	0.900	0.973	0.999	0.941	0.892	0.978	0.927	0.953	0.912	0.947
AUPRO@30%	0.980	0.966	0.982	0.944	0.924	0.960	0.980	0.982	0.970	0.981	0.967
AUPRO@10%	0.940	0.901	0.947	0.897	0.782	0.886	0.941	0.947	0.911	0.944	0.910
AUPRO@5%	0.882	0.818	0.894	0.841	0.681	0.798	0.882	0.894	0.843	0.888	0.842
AUPRO@1%	0.469	0.412	0.488	0.478	0.348	0.393	0.465	0.485	0.450	0.476	0.446
P-AUROC	0.997	0.991	0.999	0.973	0.978	0.991	0.998	0.998	0.997	0.997	0.992
cos_sim_2Dto3D	0.990	0.987	0.988	0.986	0.988	0.983	0.989	0.990	0.984	0.983	0.987
cos_sim_3Dto2D	0.815	0.881	0.828	0.786	0.881	0.821	0.806	0.837	0.851	0.855	0.836

3D到2D替换为重构	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	0.995	0.839	0.968	0.994	0.926	0.887	0.950	0.936	0.958	0.881	0.933
AUPRO@30%	0.980	0.948	0.982	0.943	0.919	0.962	0.979	0.983	0.970	0.981	0.965
AUPRO@10%	0.940	0.852	0.946	0.898	0.765	0.892	0.937	0.948	0.910	0.943	0.903
AUPRO@5%	0.882	0.751	0.893	0.843	0.662	0.810	0.875	0.897	0.842	0.886	0.834
AUPRO@1%	0.473	0.362	0.483	0.493	0.338	0.406	0.451	0.491	0.451	0.470	0.442
P-AUROC	0.997	0.984	0.998	0.972	0.976	0.992	0.998	0.998	0.997	0.996	0.991
cos_sim_2Dto3D	0.990	0.987	0.988	0.986	0.988	0.983	0.989	0.990	0.984	0.983	0.987
cos_sim_3Dto2D	0.796	0.808	0.808	0.762	0.833	0.791	0.785	0.814	0.829	0.813	0.804

3D到2D替换为重构(跳跃连接)	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	0.997	0.826	0.969	0.991	0.927	0.864	0.954	0.924	0.960	0.895	0.931
AUPRO@30%	0.979	0.954	0.982	0.942	0.922	0.965	0.979	0.983	0.970	0.981	0.966
AUPRO@10%	0.939	0.867	0.946	0.895	0.775	0.897	0.937	0.948	0.910	0.943	0.906
AUPRO@5%	0.881	0.766	0.893	0.839	0.671	0.814	0.874	0.895	0.841	0.886	0.836
AUPRO@1%	0.468	0.369	0.483	0.482	0.343	0.407	0.450	0.489	0.449	0.474	0.441
P-AUROC	0.997	0.986	0.998	0.972	0.977	0.993	0.998	0.998	0.997	0.996	0.991
cos_sim_2Dto3D	0.990	0.987	0.988	0.986	0.988	0.983	0.989	0.990	0.984	0.983	0.987
cos_sim_3Dto2D	0.815	0.854	0.832	0.781	0.869	0.810	0.798	0.835	0.850	0.843	0.829

3D到2D替换为重构(跳跃连接)更换数据集	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ
AUPRO@30%	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ
AUPRO@10%	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ
AUPRO@5%	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ
AUPRO@1%	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ
P-AUROC	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ
cos_sim_2Dto3D	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ
cos_sim_3Dto2D	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ

3D到2D替换为KAN(128)更换数据集	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	0.997	0.863	0.975	0.988	0.974	0.858	0.964	0.899	0.974	0.891	0.938
AUPRO@30%	0.979	0.956	0.982	0.941	0.945	0.961	0.980	0.982	0.973	0.981	0.968
AUPRO@10%	0.938	0.870	0.946	0.894	0.841	0.885	0.939	0.947	0.921	0.944	0.913
AUPRO@5%	0.878	0.772	0.892	0.836	0.748	0.798	0.879	0.895	0.856	0.887	0.844
AUPRO@1%	0.463	0.372	0.483	0.477	0.382	0.394	0.460	0.487	0.454	0.473	0.445
P-AUROC	0.997	0.986	0.998	0.972	0.985	0.992	0.998	0.998	0.998	0.996	0.992
cos_sim_2Dto3D	0.990	0.988	0.988	0.986	0.989	0.984	0.989	0.990	0.984	0.985	0.987
cos_sim_3Dto2D	0.820	0.839	0.834	0.795	0.874	0.807	0.813	0.834	0.882	0.862	0.836

3D到2D替换为KAN(256)更换数据集	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	0.997	0.884	0.978	0.992	0.973	0.876	0.950	0.894	0.979	0.863	0.939
AUPRO@30%	0.980	0.956	0.982	0.942	0.946	0.961	0.980	0.983	0.973	0.981	0.968
AUPRO@10%	0.939	0.873	0.946	0.894	0.843	0.887	0.940	0.948	0.921	0.944	0.914
AUPRO@5%	0.881	0.776	0.893	0.837	0.750	0.803	0.880	0.896	0.857	0.888	0.846
AUPRO@1%	0.471	0.380	0.483	0.481	0.384	0.400	0.460	0.491	0.454	0.481	0.449
P-AUROC	0.997	0.987	0.999	0.972	0.985	0.992	0.998	0.998	0.998	0.996	0.992
cos_sim_2Dto3D	0.990	0.988	0.988	0.986	0.989	0.984	0.989	0.990	0.984	0.985	0.987
cos_sim_3Dto2D	0.821	0.853	0.835	0.799	0.878	0.809	0.815	0.836	0.881	0.867	0.839

优化器分离	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ
AUPRO@30%	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ
AUPRO@10%	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ
AUPRO@5%	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ
AUPRO@1%	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ
P-AUROC	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ
cos_sim_2Dto3D	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ
cos_sim_3Dto2D	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ	ㅤ

复现50epoch	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	0.989	0.903	0.981	0.990	0.973	0.898	0.952	0.914	0.971	0.856	0.943
AUPRO@30%	0.979	0.962	0.982	0.942	0.944	0.961	0.980	0.983	0.973	0.980	0.969
AUPRO@10%	0.939	0.890	0.947	0.895	0.837	0.885	0.941	0.949	0.921	0.940	0.914
AUPRO@5%	0.879	0.803	0.894	0.840	0.740	0.798	0.883	0.897	0.857	0.881	0.847
AUPRO@1%	0.467	0.402	0.486	0.488	0.376	0.396	0.466	0.494	0.456	0.469	0.450
P-AUROC	0.997	0.989	0.999	0.972	0.985	0.992	0.998	0.998	0.998	0.996	0.992
cos_sim_2Dto3D	0.990	0.987	0.988	0.986	0.988	0.984	0.989	0.990	0.984	0.983	0.987
cos_sim_3Dto2D	0.823	0.880	0.836	0.802	0.889	0.813	0.815	0.836	0.879	0.870	0.844

两组重构250epoch	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC	0.992	0.910	0.975	0.995	0.983	0.886	0.977	0.903	0.978	0.901	0.950
AUPRO@30%	0.979	0.969	0.982	0.940	0.948	0.967	0.981	0.983	0.974	0.981	0.970
AUPRO@10%	0.939	0.910	0.946	0.894	0.849	0.902	0.943	0.948	0.922	0.942	0.920
AUPRO@5%	0.880	0.832	0.892	0.839	0.756	0.818	0.886	0.896	0.859	0.884	0.854
AUPRO@1%	0.468	0.420	0.483	0.485	0.382	0.405	0.470	0.490	0.460	0.474	0.454
P-AUROC	0.997	0.992	0.998	0.971	0.986	0.994	0.998	0.998	0.998	0.997	0.993
cos_sim_2Dto3D	0.989	0.987	0.987	0.985	0.988	0.983	0.988	0.989	0.982	0.982	0.986
cos_sim_3Dto2D	0.825	0.892	0.831	0.802	0.893	0.818	0.818	0.827	0.884	0.873	0.846

🤗 总结归纳

为什么会出现这种反直觉的现象

我同意3D到2D的映射理论上应该比2D到3D的映射更容易，因为给定3D位置确定2D位置比为2D位置"预测深度"（在某种意义上）更简单。

然而，3D特征(1152维)相对于2D特征(768维)具有更高的维度。因此，2D到3D的映射（这里指的是特征而非输入）基本上是一个扩展过程，而3D到2D的映射则是一个压缩过程。通常来说，压缩任务更具挑战性，我推测这就是为什么3D到2D的映射最终变得更加困难的原因。

特征降维的难点:

信息压缩:降维需要在保留关键信息的同时压缩数据,这需要复杂的算法来确定哪些特征是最重要的。

非线性关系:真实世界的数据通常包含复杂的非线性关系,简单的线性降维方法可能无法捕捉这些关系。

噪声干扰:降维过程中需要区分有用信息和噪声,这是一个具有挑战性的任务。

举例:主成分分析(PCA)是一种常用的线性降维方法,但它在处理非线性数据时效果不佳。为了解决这个问题,研究人员开发了更复杂的非线性降维技术,如t-SNE和自编码器。

特征升维相对容易:

信息添加:升维通常涉及添加新信息或转换现有特征,这在技术上较为简单。

灵活性:有多种方法可以进行特征升维,如多项式特征、核方法等。

过拟合风险:虽然升维可能导致过拟合,但这个问题可以通过正则化等技术来缓解。

举例:在卷积神经网络(CNN)中,卷积层通过增加特征图的数量来实现特征升维,这是一个相对直接的过程。

📝 主旨内容

改变输入

改变架构

结果

🤗 总结归纳

为什么会出现这种反直觉的现象

📎 参考文章