Daga Iska zuwa Tufafi: Ƙirƙirar Fashion na Digital 3D Na Musamman ta hanyar Zane a AR/VR

Teburin Abubuwan Ciki

1. Gabatarwa & Bayyani

Wannan aikin yana magance wani gibi mai mahimmanci a cikin ƙaddamar da ƙirƙirar fashion na dijital. Yayin da fasahohin AR/VR ke zama manyan kayan lantarki na mabukaci, kayan aikin ƙirƙirar abubuwa na 3D a cikin waɗannan wuraren nutsuwa sun kasance masu sarkakiya kuma ba za a iya samun su ga waɗanda ba ƙwararru ba. Takardar ta ba da shawarar wani sabon tsari mai cikakken tsari wanda ke baiwa masu amfani na yau da kullum damar ƙirƙirar tufafi na 3D na musamman ta hanyar tsari mai sauƙi: zane na 3D na hannu cikin 'yanci a cikin yanayin AR/VR. Babban ƙirƙira yana cikin samfurin AI mai haifarwa wanda ke fassara waɗannan zane-zane marasa daidaito, masu dacewa da mai amfani kuma ya canza su zuwa samfuran tufafi na 3D masu inganci, cikakkun bayanai waɗanda suka dace da metaverse, gwajin kama-da-kama, da bayyanar dijital.

Muhimmancin tsarin yana da bangarori biyu: yana rage matsalar fasaha ga ƙirar fashion na 3D, yana daidaitawa da yanayin sauyin mabukaci na fasahar nutsuwa, kuma yana gabatar da wani sabon tsari don ƙirƙirar abubuwa na 3D wanda ke amfani da hulɗar ɗan adam ta halitta (zane) maimakon rufaffiyar software masu sarkakiya.

2. Hanyoyi & Tsarin Fasaha

Tsarin da aka ba da shawara, mai suna DeepVRSketch+, an gina shi bisa manyan ginshiƙai uku: sabon bayanan, samfurin haifarwa mai sharadi, da dabarun horo na musamman.

2.1. Bayanan KO3DClothes

Babban matsalar cikin binciken zane-zuwa-3D shine rashin bayanan da aka haɗa (samfurin 3D + daidaitaccen zane mai amfani). Don magance wannan, marubutan sun gabatar da KO3DClothes, sabon bayanan da ke ɗauke da dubunnan nau'ikan samfuran tufafi na 3D masu inganci da daidaitattun zane-zane na 3D waɗanda masu amfani suka ƙirƙira a cikin yanayin VR. Wannan bayanan yana da mahimmanci don horar da samfurin don fahimtar taswira daga zane-zane na ɗan adam marasa ma'ana, sau da yawa masu rikitarwa, zuwa daidaitaccen lissafi na 3D.

2.2. Tsarin DeepVRSketch+

Babban samfurin haifarwa shine samfurin yaduwa mai sharadi. Ba kamar GANs na yau da kullum waɗanda zasu iya fuskantar rugujewar yanayi da rashin kwanciyar hankali na horo ba, samfuran yaduwa sun nuna babban nasara wajen samar da ingantattun sakamako iri-iri, kamar yadda samfura kamar DALL-E 2 da Stable Diffusion suka tabbatar. Samfurin yana sanya sharadi akan tsarin samarwa akan zanen 3D na shigarwa, wanda aka ɓoye shi cikin wakilcin ɓoyayye ta hanyar mai ɓoyayyen zane na musamman. Tsarin yaduwa yana maimaita cire hayani daga rarraba Gaussian bazuwar don samar da ingantaccen tufafi na 3D voxel ko gajimare wanda ya dace da niyyar zanen.

Tsarin yaduwa na gaba yana ƙara hayani ga ainihin samfurin tufafi na 3D $x_0$ sama da matakai $T$: $q(x_t | x_{t-1}) = \mathcal{N}(x_t; \sqrt{1-\beta_t} x_{t-1}, \beta_t I)$. Tsarin baya, wanda samfurin ya koya, an ayyana shi kamar haka: $p_\theta(x_{t-1} | x_t, c) = \mathcal{N}(x_{t-1}; \mu_\theta(x_t, t, c), \Sigma_\theta(x_t, t, c))$, inda $c$ shine ɗaurin zane mai sharadi.

2.3. Koyon Tsarin Karatu Mai Daidaitawa

Don ɗaukar bambance-bambancen ingancin zane daga masu amfani masu farawa, marubutan sun yi amfani da dabarun koyon tsarin karatu mai daidaitawa. An fara horar da samfurin akan zane-zane masu tsabta, daidaitattun da aka haɗa da samfuran 3D nasu. A hankali, yayin horo, ana fallasa shi ga zane-zane tare da matakan ƙara hayani da rashin cikarwa, yana kwaikwayon ainihin shigarwar daga masu amfani waɗanda ba ƙwararru ba. Wannan yana koya wa samfurin ya zama mai ƙarfi ga shubuha da rashin daidaito.

3. Sakamakon Gwaji & Kimantawa

3.1. Ma'auni na Ƙididdiga

Takardar tana kimanta samfurin da yawa akan ma'auni ta amfani da ma'auni na gama gari na sake gina 3D:

• Nisa na Chamfer (CD): Yana auna matsakaicin nisa mafi kusa tsakanin gajimaren da aka samar da gaskiya ta ƙasa. DeepVRSketch+ ya sami CD 15% ƙasa da mafi kyawun ma'auni.
• Nisa na Mai Kewayawa Duniya (EMD): Yana kimanta kamancen rarraba duniya. Samfurin da aka ba da shawara ya nuna mafi girman aiki.
• Nisa na Fréchet Point Cloud (FPD): Daidaitawar Nisan Fréchet Inception don gajimaren 3D, yana kimanta inganci da bambancin samfuran da aka samar.

3.2. Sakamako na Halitta & Nazarin Mai Amfani

A halitta, tufafin da aka samar daga DeepVRSketch+ suna nuna ƙarin lallausan gaskiya, cikakkun bayanai (kamar wrinkles da folds), da mafi kyawun bin tsarin gaba ɗaya na zanen idan aka kwatanta da ma'auni kamar Sketch2Mesh ko VR-SketchNet. An gudanar da nazarin mai amfani mai sarrafawa tare da mahalarta 50 (gauraye na masu zane da waɗanda ba masu zane ba). Mahalarta sun yi amfani da rufaffiyar zanen AR/VR don ƙirƙirar tufafi kuma sun ƙididdige tsarin. Babban binciken:

• Maki na Amfani: 4.3/5.0 don sauƙin amfani.
• Gamsuwar Fitowa: 4.1/5.0 don ingancin samfurin 3D da aka samar.
• Waɗanda ba masu zane ba sun ba da rahoton ƙaramin shinge da aka gane don shiga idan aka kwatanta da software na 3D na gargajiya kamar Blender ko CLO3D.

Hoto. 1 a cikin takardar yana taƙaita bututun a bayyane: Mai amfani yana zane a VR -> Samfurin AI yana sarrafa zane -> An samar da samfurin 3D na gaskiya -> An nuna samfurin a AR don hangen nesa/gwajin kama-da-kama.

4. Cikakken Bincike & Hasashen Kwararru

Cikakken Hasashe: Wannan takarda ba kawai game da mafi kyawun samfurin 3D ba ne; yana da cikakken tsari akan bututun ƙaddamarwa don yanar gizo mai nutsuwa. Marubutan sun gano daidai cewa babban app don mabukacin AR/VR ba kawai cinyewa ba ne, amma ƙirƙira. Ta hanyar amfani da harshen zane mai sauƙi—ƙwarewar ɗan adam ta asali—suna ƙetare babban koyon siffar polygonal, suna kai hari kai tsaye ga babban toshewar amfani don abubuwan da mai amfani ya samar na 3D. Hanyarsu tana kama da falsafar da ke bayan kayan aiki kamar Google's Quick Draw ko RunwayML, waɗanda ke ɗauke da AI mai sarkakiya zuwa rufaffiyar sauƙi.

Kwararar Hankali: Hankali yana da ban sha'awa: 1) Kayan aikin AR/VR suna zama kayayyaki (Meta Quest, Apple Vision Pro). 2) Saboda haka, babban tushen mai amfani don abubuwan nutsuwa yana fitowa. 3) Wannan yana haifar da buƙatar kadarorin dijital na musamman (fashion kasancewar babban ɗan takara). 4) Kayan aikin ƙirƙirar 3D na yanzu ba su dace da wannan babban kasuwa ba. 5) Magani: Taswira ƙwarewar ɗan adam kusan gabaɗaya (zane) akan fitowar 3D mai sarkakiya ta hanyar mai fassara AI mai ƙarfi (samfurin yaduwa). Gabatar da bayanan KO3DClothes wani muhimmin abu ne, sau da yawa ana yin watsi da shi, na abubuwan more rayuwa wanda ke ba da damar wannan fassarar, yana tunawa da yadda ImageNet ya haifar da hangen nesa na kwamfuta.

Ƙarfi & Kurakurai: Babban ƙarfi shine cikakken tsari, ƙirar mai amfani na dukan bututun, daga shigarwa (zane na VR) zuwa fitarwa (kadaran 3D mai amfani). Amfani da samfurin yaduwa mai sharadi shine na zamani kuma an ba da hujja don ɗaukar rarraba yanayi iri-iri na yiwuwar tufafi daga zane guda ɗaya. Duk da haka, kuskuren—gama gari ga yawancin takardun AI-don-ƙirƙira—yana cikin kimanta "ƙirƙira." Tsarin yana ƙware wajen fassarar da ƙaddamarwa daga zane, amma shin yana ba da damar sabon abu na gaskiya, ko kuma yana kawai dawo da haɗa alamu daga bayanan horonsa? Haɗarin shine haɗakar salon, wani rami da aka lura a cikin wasu samfuran rubutu-zuwa-hoto. Bugu da ƙari, farashin lissafi na samfuran yaduwa don ƙididdiga na ainihin lokaci a cikin saitin VR na mabukaci ba a magance shi sosai ba, yana haifar da yuwuwar shinge ga hulɗar mara tsangwama.

Hasashe Masu Aiki: Ga ƴan masana'antu, abin da za a ɗauka nan take shine saka hannun jari a cikin kayan aikin ƙirƙirar abun ciki mai sauƙi, masu ƙarfin AI a matsayin babban ɓangare na duk wani dabarun metaverse ko dandalin nutsuwa. Masu riƙe dandamali (Meta, Apple, Roblox) yakamata su kalli kayan aiki irin wannan a matsayin muhimman abubuwan SDK don tayar da tattalin arzikinsu. Ga alamar fashion, samfurin yana nuna bayyananniyar hanya don shigar da abokan ciniki a cikin haɗin gwiwar zane da keɓancewar samfurin kama-da-kama a sikeli. Hanyar binciken da za a kula shine motsi daga fitarwar voxel/gajimare zuwa nau'ikan siffa masu sauƙi, masu rai, da shirye-shiryen samarwa, mai yuwuwar haɗa simintin lissafi na lissafi don lallausan, kamar yadda ake gani a cikin aikin NVIDIA akan AI da lissafi.

5. Zurfin Fasaha

Samfurin yaduwa mai sharadi yana aiki a cikin sararin ɓoyayye da aka koya. Mai ɓoyayyen zane $E_s$ yana jefa gajimaren zane na 3D $S$ cikin vector ɓoyayye $z_s = E_s(S)$. Wannan vector mai sharadi $z_s$ ana shigar da shi cikin samfurin yaduwa na U-Net mai cire hayani a yadudduka da yawa ta hanyoyin hankali na giciye: $\text{Attention}(Q, K, V) = \text{softmax}(\frac{QK^T}{\sqrt{d}})V$, inda $Q$ wani tsinkaya ne na shigarwar hayani $x_t$, kuma $K, V$ tsinkaya ne na ɓoyayyen zane $z_s$. Wannan yana baiwa samfurin damar daidaita tsarin cire hayani tare da siffofi da fasalulluka na ma'ana na zane a matakan ƙuduri daban-daban.

Aikin asara shine ƙayyadaddun ƙayyadaddun ƙayyadaddun ƙayyadaddun ƙayyadaddun bayanai, yana mai da hankali kan hasashen hayanin da aka ƙara a kowane mataki: $L(\theta) = \mathbb{E}_{t, x_0, \epsilon} [\| \epsilon - \epsilon_\theta(x_t, t, z_s) \|^2]$, inda $\epsilon$ shine ainihin hayani kuma $\epsilon_\theta$ shine hasashen samfurin.

6. Tsarin Bincike & Nazarin Hali

Tsarin don Kimanta Kayan Aikin AI na Ƙirƙira:

1. Samun dama: Yanayin shigarwa na halitta (misali, zane da code).
2. Aminci: Ingancin fitarwa da bin niyya (wanda aka auna ta CD, EMD, nazarin mai amfani).
3. Sarrafawa: Girman sarrafa mai amfani akan fitarwa (siffar duniya da cikakkun bayanai na gida).
4. Gama gari: Ikon sarrafa shigarwar mai amfani iri-iri, da ba a gani ba da salo.
5. Shirye-shiryen Samarwa: Daidaitaccen tsarin fitarwa (misali, .obj, .fbx, taswirorin UV).

Nazarin Hali: Ƙirar "Rigar Draped Asymmetric"

1. Aikin Mai Amfani: A cikin VR, mai amfani yana zane siffar riga tare da babban abin wuya a kafada ɗaya da kuma layin hem mai gudana, mara daidaito.
2. Sarrafa Tsarin: Mai ɓoyayyen zane yana ɗaukar siffar asymmetric na duniya da niyya ta gida don lallausan. Samfurin yaduwa, wanda aka sanya shi a kan wannan, ya fara cire hayani. Koyon tsarin karatu yana tabbatar da cewa ko da zanen yana da sako-sako, samfurin yana haɗa layukan gudana da lissafin lallausan yadi mai laushi.
3. Fitarwa: Tsarin yana samar da siffar 3D na riga. An fahimci babban abin wuya a matsayin ninki mai tsari, yayin da layin hem yana da wrinkles iri-iri, masu kama da na halitta. Mai amfani zai iya juyawa, duba a AR akan avatar na kama-da-kama, kuma zaɓi ya inganta ta hanyar sake zane a wurare.
4. Kimantawa ta hanyar Tsarin: Babba akan Samun dama da Gama gari (ya sarrafa ƙira mara al'ada). Aminci yana da matsayi mai girma. Sarrafawa matsakaici ne—mai amfani ba zai iya sauƙaƙe ainihin adadin wrinkles bayan samarwa ba, yana nuna wani yanki na bincike na gaba.

7. Aikace-aikace na Gaba & Hanyoyi

• Haɗin gwiwar Ƙirƙira na Ainihin Lokaci & Zanen Zamantakewa: Masu amfani da yawa a cikin sararin VR ɗaya suna zane da maimaitawa akan tufafi ɗaya lokaci ɗaya, tare da samfuran da AI ke samarwa a rayuwa.
• Haɗawa tare da Simintin Lissafi: Haɗa samfurin haifarwa tare da na'urori na kwaikwayon tufafi na ainihin lokaci (misali, bisa NVIDIA FleX ko PyBullet) don tabbatar da cewa tufafin da aka samar suna motsawa da lallausan a kan avatar masu rai tun daga farko.
• Gyaran Jagora ta Rubutu & Murya: Yanayi mai yawa. misali, "Sanya hannun riga ya yi ƙura" ta hanyar umarnin murya ko rubutu, yana inganta fitowar tushen zane, kama da InstructPix2Pix.
• Gada Kai tsaye-zuwa-Samar-Dijital: Don fashion na zahiri, faɗaɗa bututun don samar da alamu na ɗinki 2D daga samfurin 3D, yana taimakawa wajen ƙirƙirar tufafi na zahiri.
• Mataimakin Fashion na AI Na Musamman: Wakilin AI wanda ke koyon salon mutum na mai amfani daga tarihin zanensu kuma zai iya ba da shawarar gyare-gyare, cika ɓangarorin zane, ko samar da sababbin ra'ayoyi masu dacewa da ɗanɗanonsu.

8. Nassoshi

1. Zang, Y., Hu, Y., Chen, X., da sauransu. "Daga Iska zuwa Tufafi: Fashion na Dijital 3D Na Musamman tare da Zanen 3D Mai Nutsuwa na AR/VR." Jaridar Fayilolin Latex, 2021.
2. Ho, J., Jain, A., & Abbeel, P. "Samfuran Ƙididdiga na Yaduwa na Cire Hayani." Ci gaba a cikin Tsarin Bayanai na Neural (NeurIPS), 2020. (Takarda mai mahimmanci ta samfurin yaduwa).
3. Rombach, R., Blattmann, A., Lorenz, D., da sauransu. "Haɗin Hoton Babban Ƙuduri tare da Samfuran Yaduwa na ɓoyayye." Gudanar da Taron IEEE/CVF akan Hangen Nesa na Kwamfuta da Tsarin Alamu (CVPR), 2022. (Akan yaduwar sararin ɓoyayye).
4. Isola, P., Zhu, J., Zhou, T., & Efros, A. A. "Fassarar Hoto-zuwa-Hoto tare da Cibiyoyin Adawa masu Sharadi." CVPR, 2017. (Tsarin Pix2Pix, tushen don samarwa mai sharadi).
5. NVIDIA. "NVIDIA Cloth & Simintin Lissafi." https://www.nvidia.com/en-us/design-visualization/technologies/cloth-physics-simulation/
6. Meta. "Dandalin Kasancewar: Insight SDK don Bin Hannu." https://developer.oculus.com/documentation/unity/ps-hand-tracking/ (Mai dacewa don yanayin shigarwa).