1. Gabatarwa
Daidaitawar hotunan fuskokin mutum aiki ne muhimmi a cikin daukar hoto na lissafi da gyaran hoto, wanda ke nufin haɗa abin da ke gaba cikin sabon bayan gida cikin sauƙi. Hanyoyin gargajiya sau da yawa sun kasa yin la'akari da haduwar haske masu sarkakiya, wanda ke haifar da sakamako marar gaskiya. Wannan takarda ta gabatar da Daidaitawar Haske Mai Kyau, sabon tsarin tushen yaduwa wanda ke ƙirƙira da canja yanayin haske daga bayan gida zuwa hoton fuskar da ke gaba, yana cimma ingantaccen kamannin hoto.
2. Hanyar Aiki
Tsarin da aka gabatar yana aiki a cikin matakai uku na asali, yana wucewa da sauƙaƙan daidaitawar launi don cimma haɗin kai na haske na gaskiya.
2.1 Tsarin Wakilcin Haske
Wannan tsari yana ciro bayanan haske na ɓoye (misali, shugabanci, ƙarfi, zafin launi) daga hoton bayan gida guda ɗaya. Yana ɓoye waɗannan bayanan cikin wakilcin haske na ɓoye $L_{bg}$ wanda ke aiki azaman siginar sharadi don tsarin yaduwa. Wannan yana ƙetare buƙatar taswirorin muhalli na HDR bayyananne yayin ƙididdigewa.
2.2 Cibiyar Sadarwar Daidaitawa
Don kafa fasalin haske da aka koya a cikin sarari mai ma'ana ta zahiri, an gabatar da cibiyar sadarwar daidaitawa. Tana daidaita fasalin haske da aka samo daga hoto $L_{bg}$ da fasalin da aka ciro daga cikakken taswirorin muhalli na panorama $L_{env}$ yayin horo. Wannan haɗin yana tabbatar da cewa samfurin ya koyi ingantaccen fahimtar haske na wuri wanda zai iya yaduwa, kamar yadda aka tabbatar da shi ta bayanan kamar Laval Indoor HDR.
2.3 Hanyar Bayanan Ƙirƙira
Wani sabon abu mai mahimmanci shine hanyar kwaikwayon bayanai wanda ke samar da nau'ikan nau'ikan horo masu inganci. Yana haɗa batutuwan mutane daga cikin bayanan da suka wanzu (misali, FFHQ) akan bayanan gida daban-daban tare da sanannen haske, yana ƙirƙira bayanan da aka haɗa {gaba, bayan gida, gaskiyar daidaitawa} ba tare da buƙatar ɗaukar haske mai tsada ba. Wannan yana magance babban matsalar bayanai a fagen.
3. Cikakkun Bayanan Fasaha
Samfurin ya ginu akan samfurin yaduwa na ɓoye da aka riga aka horar (LDM). Tsarin samarwa na asali yana jagorantar ta yanayin haske. Tsarin kawar da hayaniya a lokacin $t$ ana iya tsara shi kamar haka:
$\epsilon_\theta(z_t, t, \tau(L_{bg}), \tau(mask))$
inda $z_t$ shine ɓoyayyen ɓoyayye, $\epsilon_\theta$ shine mai kawar da hayaniya na UNet, $\tau(\cdot)$ yana nuna masu ɓoyayyen sharadi, $L_{bg}$ shine wakilcin haske na bayan gida, kuma $mask$ shine abin rufe fuska na gaba. Cibiyar sadarwar daidaitawa tana inganta asarar daidaitawar fasali $\mathcal{L}_{align} = ||\phi(L_{bg}) - \psi(L_{env})||_2$, inda $\phi$ da $\psi$ su ne cibiyoyin sadarwar tsinkaya.
4. Gwaje-gwaje & Sakamako
An kimanta hanyar a kan mafi kyawun daidaitawa (misali, DoveNet, S2AM) da ma'auni na sake haskakawa. Ma'auni na ƙididdiga (PSNR, SSIM, LPIPS, FID) da nazarin masu amfani sun ci gaba da sanya Daidaitawar Haske Mai Kyau a matsayi mafi girma don kamannin gani da daidaiton haske.
Nazarin Hoto na 1: Hoto na 1 na takardar yana nuna ƙarfin iyawar samfurin. Yana nuna misalai huɗu na zahiri inda haɗin kai kai tsaye (abin da aka liƙa akan bayan gida) yayi kama da rashin daidaiton shugabanci na haske da sanya inuwa. Akasin haka, sakamakon samfurin yana sake haskaka abin da ke gaba da gaskiya: launukan fata sun dace da launin muhalli, an sake sanya haske da inuwa don dacewa da sabon tushen haske, kuma gabaɗayan haɗin yana bayyana kamar hoto na gaskiya.
5. Tsarin Nazari: Fahimta ta Asali & Zargi
Fahimta ta Asali: Babban nasarar takardar ita ce gane cewa daidaitawa ta gaskiya matsala ce ta sake haskakawa a ɓoye. Yayin da aikin da ya gabata kamar CycleGAN (Zhu et al., 2017) ya yi fice a canja salon da ba a haɗa shi ba, ya ɗauki haske a matsayin salon launi kawai. Wannan aikin ya gano daidai shugabanci na haske, jefa inuwa, da fitattun haske a matsayin abubuwan lissafi da na zahiri waɗanda dole ne a ƙirƙira su a bayyane, ba kawai a daidaita su ta ƙididdiga ba. Yana amfani da hikimar tsarin samfuran yaduwa da wayo don magance wannan matsala ta juyawa mara kyau.
Kwararar Ma'ana: Hanyar aiki ta matakai uku tana da ma'ana sosai. 1) Gane haske daga hoto (matsala mai wahala). 2) Kafa wannan fahimta a cikin sanannen, cikakken wakilci (taswirorin panorama) yayin horo don tabbatar da yiwuwar zahiri. 3) Ƙirƙira ɗimbin bayanan horo don koya wa samfurin wannan taswira mai sarkakiya. Dabarar bincike ce ta "ayyana, daidaita, ƙidaya" wanda aka aiwatar da kyau.
Ƙarfi & Kurakurai: Babban ƙarfinsa shine amfaninsa na aikace-aikace—yana aiki tare da hoton bayan gida guda ɗaya, babbar fa'ida akan hanyoyin da ke buƙatar panorama na HDR. Hanyar bayanan ƙirƙira babban nasara ce don haɓakawa. Duk da haka, aibi yana cikin duhunsa: a matsayin samfurin yaduwa mai yawa, akwatin baƙar fata ne. Ba mu sami samfurin haske mai fassara (misali, vector coefficient na SH 3D) a matsayin sakamako ba, yana iyakance amfaninsa a cikin hanyoyin zane na gaba. Hakanan yana iya fuskantar wahala tare da bambance-bambancen haske mai tsanani ko kayan haske masu haske, yanayin gazawar gama gari ga samfuran samarwa.
Fahimta Mai Aiki: Ga ƙungiyoyin samfur, wannan API ce da za a iya haɗawa don kayan aikin gyaran hoto masu inganci. Ga masu bincike, makomar a bayyane take: 1) Rarraba lambar haske ta ɓoye zuwa sigogi masu fassara (shugabanci, ƙarfi, laushi). 2) Ƙara zuwa bidiyo don daidaiton lokaci—kalubale mai girma amma wajibi. 3) Haɗin kai tare da al'ummar NeRF/Ƙirƙirar 3D. Ƙarshen ma'ana ba kawai daidaita Layer 2D ba ne, amma saka kayan 3D da aka sake haskaka cikin wuri, hangen nesa da ayyuka daga MIT CSAIL da Google Research suka raba.
6. Aikace-aikace na Gaba & Jagorori
- Ƙarfafa & Gaskiyar Zamani: Daidaitawar cikakken lokaci na ciyarwar kyamara mai rai tare da muhallin ƙira don gogewa mai nutsewa.
- Fim & Bayan Samar da Bidiyo: Gyaran haske mai daidaitawa da kai tsaye don haruffan da aka haɗa cikin bayanan gida na CGI, yana rage farashin VFX sosai.
- Gwaji na Ƙira & Kayan Kwalliya: Yin amfani da haske da inuwa na gaskiya ga samfura ko tufafi da aka haɗa akan hotunan mai amfani.
- Kasancewar Nesa & Taron Bidiyo: Daidaita yanayin haske ga duk mahalarta don ƙirƙirar wurin taro na ƙira mai haɗin kai.
- Shugabanci na Bincike: Haɗawa tare da samfuran samarwa masu sanin 3D (misali, 3D Gaussian Splatting) don cimma daidaiton ra'ayi na sake haskakawa da jefa inuwa.
7. Nassoshi
- Ren, M., Xiong, W., Yoon, J. S., et al. (2024). Daidaitawar Haske Mai Kyau: Sauyin Bayanan Fuskokin Mutum Tare da Lura da Haske. arXiv:2312.06886v2.
- Zhu, J., Park, T., Isola, P., & Efros, A. A. (2017). Fassarar Hotuna zuwa Hotuna marasa Haɗawa ta amfani da Cibiyoyin Sadarwar Adawa masu Daidaitaccen Zagaye. IEEE ICCV.
- Rombach, R., Blattmann, A., Lorenz, D., Esser, P., & Ommer, B. (2022). Haɗin Hotuna Mai Ƙarfi tare da Samfuran Yaduwa na ɓoye. IEEE CVPR.
- Zhang, L., et al. (2021). S2AM: Cibiyar Sadarwa Mai Sauƙi don Daidaitawar Hotuna. ACM MM.
- Debevec, P. (2012). Matakin Haske da Aikace-aikacensa ga ƴan wasan dijital na Hotuna na Gaskiya. Kwasa-kwasan SIGGRAPH.
- Mildenhall, B., et al. (2020). NeRF: Wakilcin Wurare azaman Filayen Haske na Jijiya don Haɗin Duba. ECCV.