Hasken cikin Motsi (LIMO) yana gabatar da sabuwar hanya mai tushen yaduwa don kimanta haske mai girma mai ƙarfi (HDR) na lokaci da sarari daga bidiyo guda ɗaya. Babban ƙalubalen da aka magance shi ne shigar da abubuwa ko ƴan wasan kwaikwayo na kwaikwayo cikin hotunan aiki na gaske, aiki mai mahimmanci a cikin samar da bidiyo na kwaikwayo, ƙarin gaskiya, da tasirin gani. Hanyoyin gargajiya sun dogara da binciken haske na zahiri, waɗanda ke tsoma baki kuma ba su da amfani ga yawancin yanayi. LIMO yana sarrafa wannan ta hanyar kimanta hasken da ke tushen sarari (ya bambanta da matsayi na 3D), daidaitaccen lokaci (ya daidaita akan lokaci), kuma yana ɗaukar cikakken kewayon HDR daga haske kaɗan kai tsaye zuwa tushe masu haske kai tsaye, a ciki da waje.
Takardar ta tabbatar cewa dabarar kimanta haske gabaɗaya dole ne ta cika iyawa guda biyar: 1) Tushen sarari a takamaiman wuri na 3D, 2) Daidaitawa ga bambancin lokaci, 3) Hasashen haske na HDR daidai, 4) Sarrafa duka tushen haske na kusa (ciki) da nesa (waje), da 5) Kimanta rarraba haske mai yuwuwa tare da cikakken bayani mai sauri. An sanya LIMO a matsayin tsari na farko da aka haɗa wanda ke nufin duka biyar.
Shigarwa: Hoton guda ɗaya ko jerin bidiyo da matsayi na 3D da ake nufi. Tsari: 1) Yi amfani da mai kimanta zurfi na guda ɗaya (misali, [5]) don samun zurfin kowane pixel. 2) Lissafa sabbin taswirorin yanayin lissafi daga zurfi da matsayin da ake nufi. 3) Sanya ƙirar yaduwa da aka gyara tare da waɗannan taswirorin don samar da hasashen madubi da siffofi masu yaduwa a bayyanawa da yawa. 4) Haɗa waɗannan hasashen cikin taswirar muhalli ta ƙarshe ta HDR.
Marubutan sun gano cewa zurfin kawai yana ba da wakilcin fage mara cikawa don haske na gida. Sun gabatar da ƙarin yanayin lissafi wanda ke ɓoye matsayin dangi na lissafin fage zuwa wurin da ake nufi. Wannan yana iya haɗawa da wakiltar vectors ko filayen nisa da aka sanya hannu daga wurin da ake nufi zuwa saman da ke kewaye, yana ba da mahimman bayanai don rufewa da kusancin tushen haske waɗanda taswirorin zurfi kawai ba su da.
LIMO ya ginu akan ƙirar yaduwa a ɓoye da aka riga aka horar (misali, Stable Diffusion). An gyara shi akan babban bayanan da aka keɓance na fage na ciki da waje, kowannensu yana haɗe da binciken haske na HDR da aka daidaita a lokaci da sarari da aka ɗauka a wurare daban-daban. An canza shigar da yanayin don karɓar taswirorin lissafi (zurfi + matsayi na dangi) tare da hoton RGB. An horar da ƙirar don rage hayaniyar ko dai taswirar madubi na madubi ko taswirar haske mai yaduwa a takamaiman matakin bayyanawa.
Horon yana iya haɗawa da aikin asara wanda ya haɗa asarar fahimta (misali, LPIPS) don bayani da asarar L1/L2 don daidaiton haske, kama da hanyoyin a cikin ayyukan fassarar hoto zuwa hoto kamar waɗanda Isola et al. a cikin Pix2Pix suka fara.
Babban ƙirar fasaha don sake gina HDR yana kwance a cikin hasashen bayyanawa da yawa da haɗawa. Bari $I_{m}^{e}(x)$ da $I_{d}^{e}(x)$ su wakilci hotunan madubi da masu yaduwa da aka hasashe a bayyanawa $e$ don matsayi da ake nufi $x$. Taswirar muhalli ta ƙarshe ta HDR $L_{env}(\omega)$ an sake gina ta ta hanyar warware matsalar ingantawa ta hanyar zane mai banbanci:
$$ L_{env} = \arg\min_{L} \sum_{e} \left\| R(L, e) - \{I_{m}^{e}, I_{d}^{e}\} \right\|^2 $$
Inda $R(L, e)$ mai zane ne mai banbanci wanda ke kwaikwayon hoton da aka kafa akan madubi/siffa mai yaduwa ta taswirar muhalli $L$ a bayyanawa $e$. Wannan yana tabbatar da daidaiton zahiri a cikin bayyanawa da nau'ikan siffa.
Da alama takardar tana ƙima ta amfani da ma'auni na yau da kullun don kimanta haske da haɗin sabon ra'ayi:
An yi iƙirarin cewa LIMO ya kafa sakamako na zamani a cikin duka daidaiton sarrafa sarari da amincin hasashe idan aka kwatanta da ayyukan da suka gabata kamar [15, 23, 25, 26, 28, 30, 35, 41, 50].
Hoto na 1 a cikin PDF yana nuna mahimman sakamako: 1) Daidaiton tushen sarari: Abu na kwaikwayo yana nuna inuwa da inuwa daidai lokacin da aka sanya shi a wurare daban-daban a cikin daki. 2) Daidaiton lokaci: Hasken akan abu na kwaikwayo yana canzawa da gaske yayin da kyamara ke motsawa. 3) Aikace-aikacen samarwa na kwaikwayo: An ɗauki ɗan wasan kwaikwayo a cikin matakin haske an haɗa shi cikin gaskiya cikin fage na gaske ta amfani da hasken da aka kimanta na LIMO, yana nuna madubi na gaske da haɗawa.
Sakamakon ya nuna cewa LIMO yayi nasarar hasashen cikakkun bayanai masu sauri (misali, firam ɗin taga, madubi masu rikitarwa) da kewayon motsi mai faɗi (misali, hasken rana mai haske da kusurwoyi masu duhu).
Nazarin cirewa zai tabbatar da mahimman zaɓin ƙira: 1) Tasirin sabon yanayin lissafi: Nuna cewa ƙirar da aka sanya kawai akan zurfi suna samar da haske mai tushen sarari mara daidai. 2) Hasashen bayyanawa da yawa da na guda ɗaya: Nuna wajibcin bututun bayyanawa da yawa don dawo da cikakken kewayon HDR. 3) Ƙirar yaduwa da aka gabatar: Kwatanta gyara babban ƙirar tushe da horar da cibiyar sadarwa ta musamman daga farko.
Mahimmin Bayani: Babban nasarar LIMO ba wai kawai wani ci gaba ne a cikin daidaiton kimanta haske ba. Yana da ma'anar daga fahimtar fage na duniya zuwa mahallin haske na aiki, na gida. Yayin da hanyoyin da suka gabata kamar Gardner et al. [15] ko Srinivasan et al. [41] suka ɗauki haske a matsayin kadarorin fage, LIMO ya gane cewa don shigarwa mai amfani, hasken a takamaiman voxel inda abun ku na CG yake shine duk abin da ke da mahimmanci. Wannan yana canza tsarin daga "Menene hasken wannan ɗakin?" zuwa "Menene hasken nan?" – tambaya mafi daraja ga bututun VFX.
Tsarin Ma'ana: Gine-ginen fasaha yana da kyau sosai. Maimakon tilasta cibiyar sadarwa guda ɗaya ta fitar da taswira mai rikitarwa, mai girma mai ƙarfi (HDR) kai tsaye—aikin koma baya mai wahala—LIMO yana rarraba matsalar. Yana amfani da ƙirar samarwa mai ƙarfi (yaduwa) a matsayin "mai hasashen bayani," wanda aka sanya shi akan alamun lissafi masu sauƙi, don samar da lura na wakili (hotunan siffa). Wani mataki na haɗawa na daban, na zahiri (zane mai banbanci) daga nan yana warware filin haske na asali. Wannan rabuwar "abin da aka gabatar na ilmantarwa" da "ƙuntatawa na kimiyyar lissafi" tsari ne mai ƙarfi, mai tunawa da yadda NeRF ya haɗu da filayen haske da aka koya tare da lissafin zanen girma.
Ƙarfi & Kurakurai: Babban ƙarfin shi ne burinsa na gaba ɗaya. Magance duka iyawa guda biyar a cikin ƙira ɗaya motsi ne mai ƙarfi wanda, idan ya yi nasara, yana rage rikitarwar bututu sosai. Amfani da abubuwan da aka gabatar na yaduwa don cikakken bayani mai sauri shima yana da hankali, yana amfani da biliyoyin daloli na saka hannun jari na al'umma a cikin ƙirar tushe. Duk da haka, kuskure mai mahimmanci yana kwance a cikin jerin dogaro. Ingancin yanayin lissafi (zurfi + matsayi na dangi) yana da mahimmanci. Kurakurai a cikin kimanta zurfi na guda ɗaya—musamman ga saman da ba na Lambertian ko na bayyane—za su watsa kai tsaye cikin hasashen haske mara daidai. Bugu da ƙari, aikin hanyar a cikin fage masu ƙarfi sosai tare da tushen haske masu saurin motsi ko canje-canjen haske masu tsauri (misali, canjin haske yana jujjuyawa) ya kasance tambaya a buɗe, kamar yadda tsarin yanayin lokaci ba a yi cikakken bayani ba.
Bayanai Masu Aiki: Ga ɗakunan VFX da ƙungiyoyin samarwa na kwaikwayo, abin da za a yi nan da nan shi ne gwada matsin lamba na tushen sarari. Kar a kawai kimanta akan hotuna masu tsayi; matsar da abu na kwaikwayo tare da hanya kuma duba don flickering ko canje-canjen haske mara dabi'a. Dogaro da kimanta zurfi yana nuna hanyar haɗakarwa: amfani da LIMO don kimanta farko, amma barin masu fasaha su inganta sakamakon ta amfani da ma'auni kaɗan, sauƙin ɗaukar ma'auni na zahiri (misali, ƙwallon chrome guda ɗaya da aka harba a kan saiti) don gyara kurakurai na tsarin. Ga masu bincike, mataki na gaba a bayyane shi ne rufe tazarar yanki. Bayanan da aka gyara suna da mahimmanci. Haɗin gwiwa tare da ɗakunan studio don ƙirƙirar babban bayanan, bayanai daban-daban na ɗaukar fage/LiDAR/binciken haske na zahiri—kamar yadda Waymo ya yi don tuƙi mai cin gashin kansa—zai zama mai canza wasa, yana motsa fagen bayan bayanan roba ko iyakantaccen bayanin gaske.