Many　content-based　3D　model　retrieval　techniques　have　been　proposed　in　the　last　decade.　Compared　with　other　methods,　light　field　descriptor　(LFD)　is　generally　acknowledged　as　the　best　model　representation.　Taking　LFD　as　example,　we　analyze　the　disadvantage　of　the　model　descriptors　based　on　visual　similarity　and　propose　a　modified　descriptor　which　preserves　the　exact　geometrical　features　of　the　model　in　each　view.　Since　the　modification　avoids　distortion,　the　visual　differences　between　dissimilar　models　are　kept.　The　experimental　results　show　that　the　improved　method　outperforms　the　original　LFD　approach　significantly.　According　to　our　analysis　and　the　results,　we　draw　a　conclusion　about　the　methods　based　on　2D　visual　similarity.　Furthermore,　a　new　user-friendly　query　interface　is　provided　to　assist　user　to　draw　2D　query　from　ordinary　picture.　This　method　not　only　makes　retrieval　process　efficient　but　also　gets　more　accurate　results　than　other　methods　proposed　before.　©　2008　IEEE.