3D　hand　pose　estimation　aims　to　regress　3D　location　of　the　hand　joints　based　on　the　input　2D　image,　which　has　been　widely　applied　to　virtual　reality,　natural　human-computer　interaction,　automatic　driving　and　other　fields.The　hand　pose　estimation　from　a　single　RGB　image　is　more　pervasive　but　faces　many　challenges.To　model　complex　hand　joints,　this　paper　proposes　a　kind　of　3D　hand　pose　estimation　method　based　on　double　branches.In　the　proposed　double　branches　network　structure,　one　is　used　to　describe　the　classical　physical　connection　among　the　joints　of　the　same　finger,　and　the　other　is　used　to　describe　symmetrical　connection　among　the　corresponding　joints　of　different　fingers.The　complementary　topology　effectively　models　complex　associations　among　the　joints.With　regard　to　each　branch,　this　paper　presents　a　kind　of　multi-scale　attention　for　hand　pose　regression　which　improves　the　hand　pose　estimation　accuracy　via　multi-scale　representation　and　scales　attention.Experimental　results　on　STB　and　FreiHand　datasets　show　that　the　proposed　method　is　superior　to　existing　hand　pose　estimation　methods,　and　the　average　joint　error　is　respectively　reduced　by　0.6　mm　and　0.8　mm　relative　to　the　baseline.　©　2023　Science　Press.　All　rights　reserved.