One　key　challenge　in　object　pose　estimation　is　learn　small　object　features　under　poor　quality　conditions　while　preventing　false　positives　for　object　boundaries.　The　current　study　rarely　focuses　on　small　objects,　and　emphasizes　specificity　while　ignores　sensitivity.　In　this　paper,　we　propose　a　multi-stage　learning　framework-SO-PERM　to　address　these　problems,　which　includes　multiple　modules　like　sub-pixel　detection,　3D　view　reconstruction　and　multi-scale　unsupervised　network.　Instead　of　using　single　object　features,　we　choose　to　combine　the　noise　distribution　and　boundary　artifacts　around　the　interference　region,　then　refine　the　sub-pixels　to　clearly　extract　the　target　boundary.　The　3D　view　reconstruction　allows　us　to　correlate　the　constructed　2D　features　with　the　real　scene.　The　further　obtained　3D　heat　map　of　simulated　objects　is　not　negligible　for　the　pose　estimation　of　tiny　targets.　In　addition,　considering　the　feature　sensitivity　of　small　targets,　we　propose　a　multi-scale　unsupervised　network　and　an　iterative　refinement　scheme　to　improve　performance,　which　combines　global　and　local　methods　to　quickly　locate　objects　and　extract　specific　features　respectively.　Extensive　experiments　on　four　benchmark　sets　not　only　demonstrate　the　robustness　of　SO-PERM,　but　also　the　feasibility　of　low-dimensional　features　to　extend　the　idea　of　high-dimensional　features.