Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
NIPS2017reading_3Dreconstruction
Search
望月紅葉さんと幸せな家庭を築きたい
January 27, 2018
Research
0
1.5k
NIPS2017reading_3Dreconstruction
望月紅葉さんと幸せな家庭を築きたい
January 27, 2018
Tweet
Share
More Decks by 望月紅葉さんと幸せな家庭を築きたい
See All by 望月紅葉さんと幸せな家庭を築きたい
shadow-detection-with-conditional-generative-adversarial-networks
momijifullmoon
0
150
unsupervised-learning-of-depth-and-ego-motion-from-monocular-video-using-3d-geometric-constraints
momijifullmoon
0
430
ABEJA Innovation Meetup NIPS PointNet++
momijifullmoon
1
490
Other Decks in Research
See All in Research
世界の人気アプリ100個を分析して見えたペイウォール設計の心得
akihiro_kokubo
PRO
62
30k
日本語新聞記事を用いた大規模言語モデルの暗記定量化 / LLMC2025
upura
0
230
Agentic AIとMCPを利用したサービス作成入門
mickey_kubo
0
630
SSII2025 [TS3] 医工連携における画像情報学研究
ssii
PRO
3
1.3k
EarthDial: Turning Multi-sensory Earth Observations to Interactive Dialogues
satai
3
210
診断前の病歴テキストを対象としたLLMによるエンティティリンキング精度検証
hagino3000
1
150
SegEarth-OV: Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images
satai
3
290
【輪講資料】Moshi: a speech-text foundation model for real-time dialogue
hpprc
3
720
大規模な2値整数計画問題に対する 効率的な重み付き局所探索法
mickey_kubo
1
380
EcoWikiRS: Learning Ecological Representation of Satellite Images from Weak Supervision with Species Observation and Wikipedia
satai
3
240
Unsupervised Domain Adaptation Architecture Search with Self-Training for Land Cover Mapping
satai
3
190
RHO-1: Not All Tokens Are What You Need
sansan_randd
1
190
Featured
See All Featured
Connecting the Dots Between Site Speed, User Experience & Your Business [WebExpo 2025]
tammyeverts
9
580
Practical Tips for Bootstrapping Information Extraction Pipelines
honnibal
PRO
23
1.5k
Evolution of real-time – Irina Nazarova, EuRuKo, 2024
irinanazarova
9
960
Designing for humans not robots
tammielis
254
25k
Raft: Consensus for Rubyists
vanstee
139
7.1k
実際に使うSQLの書き方 徹底解説 / pgcon21j-tutorial
soudai
PRO
188
55k
A Modern Web Designer's Workflow
chriscoyier
697
190k
RailsConf & Balkan Ruby 2019: The Past, Present, and Future of Rails at GitHub
eileencodes
140
34k
Easily Structure & Communicate Ideas using Wireframe
afnizarnur
194
16k
Fashionably flexible responsive web design (full day workshop)
malarkey
407
66k
Agile that works and the tools we love
rasmusluckow
331
21k
GraphQLとの向き合い方2022年版
quramy
49
14k
Transcript
̏࣍ݩ෮ݩʹؔͯ͠ Learning a Multi-View Stereo Machine NIPS2017จಡΈձˏΫοΫύου 1 ಛʹදه͕ͳ͍ݶΓɺҎԼͷࢿྉ͔ΒҾ༻ https://arxiv.org/pdf/1708.05375.pdf
Learning a Multi-View Stereo Machine ▸ චऀ • Abhishek Kar,
Christian Häne, Jitendra Malik ʢUC Berkeley) ▸ ֓ཁ • Multi View StereoʢMVSʣʹΑΔີͳ3࣍ݩ෮ݩΛDeep LearningͰEnd2Endʹֶश • MVSΛ”ֶशͰ͖Δ”ͷͰແ͍͔ͱ͍͏ٙʹ͑Δ 2
എܠ ▸ Multi View Stereoͱ 1. ಛநग़ 2. Ϛονϯά 3.
̏࣍ݩ෮ݩ 4. Τϥʔͷআڈ 3
എܠ ▸ Multi View Stereoͱ 1. ಛநग़ 2. Ϛονϯά 3.
̏࣍ݩ෮ݩ 4. Τϥʔͷআڈ ==> DeepԿͰશͯղܾͰ͖ͦ͏ 4
എܠ ▸ Multi View Stereoͱ 1. ಛநग़ɹ← CNNͰ͍͚Δ 2. Ϛονϯά
3. ̏࣍ݩ෮ݩ 4. Τϥʔͷআڈ 5
എܠ ▸ Multi View Stereoͱ 1. ಛநग़ 2. Ϛονϯάɹ← CNNͱRNNͰ͍͚Δ
3. ̏࣍ݩ෮ݩ 4. Τϥʔͷআڈ 6
എܠ ▸ Multi View Stereoͱ 1. ಛநग़ 2. Ϛονϯά 3.
̏࣍ݩ෮ݩɹ← DeconvͰ͍͚Δ 4. Τϥʔͷআڈ 7
എܠ ▸ Multi View Stereoͱ 1. ಛநग़ 2. Ϛονϯά 3.
̏࣍ݩ෮ݩ 4. Τϥʔͷআڈɹ← Encoder-DecoderͰ͍͚Δ 8
DeepԿͰࡾ࣍ݩ෮ݩ ▸ 3DR2N2(ECCV2016) • ෳը૾ΛΤϯίʔυ͠ɺLSTMͰϚονϯά 9 http://3d-r2n2.stanford.edu
DeepԿͰࡾ࣍ݩ෮ݩ ▸ 3D Shape Reconstruction by Modeling 2.5D Sketch (NIPS2017)
• ϦΞϧͷը૾͔Β2.5DͷεέονΛى͜͠ɺ2.5DεέονΛͱʹ 3DshapeਪఆΛEnd2EndֶशͰ͢Δ 10 https://arxiv.org/pdf/1711.03129.pdf
͢༰ ▸ શମ૾ ▸ ख๏ ▸ ࣮ݧ ▸ ·ͱΊ 11
શମ૾ 12 http://bair.berkeley.edu/blog/2017/09/05/unified-3d/
શମ૾ 13 Learnt Stereo Machines
ख๏ ▸ Image Encoder • Encoder-DecoderܕʢU-netʣͷઃܭ • Ϛονϯάʹ༻͍Δ̎DͷಛϚοϓ࡞ • ࣍ݩ2DnಛϚο
14
ख๏ ▸ Unplojection ▸ 2࣍ݩͷಛϚοϓ3࣍ݩͷຊདྷ͋Δ͖ಛϚοϓ͔ΒࣹӨ ▸ 3࣍ݩάϦουʹٯࣹӨ 15 http://bair.berkeley.edu/blog/2017/09/05/unified-3d/
ख๏ ▸ Unplojection ▸ 2࣍ݩͷಛϚοϓ3࣍ݩͷຊདྷ͋Δ͖ಛϚοϓ͔ΒࣹӨ ▸ 3࣍ݩάϦουʹٯࣹӨ 16 http://bair.berkeley.edu/blog/2017/09/05/unified-3d/
ख๏ ▸ Unplohection ▸ 2࣍ݩͷಛϚοϓ3࣍ݩͷຊདྷ͋Δ͖ಛϚοϓ͔ΒࣹӨ ▸ 3࣍ݩάϦουʹٯࣹӨ 17 http://bair.berkeley.edu/blog/2017/09/05/unified-3d/
ख๏ ▸ Unplohection ▸ 2࣍ݩͷಛϚοϓ3࣍ݩͷຊདྷ͋Δ͖ಛϚοϓ͔ΒࣹӨ ▸ 3࣍ݩάϦουʹٯࣹӨ 18 http://bair.berkeley.edu/blog/2017/09/05/unified-3d/
ख๏ ▸ Recurrent Grid Fusion • 3࣍ݩͷಛϚοϓͷϚονϯάΛGated Recurrent Unit(GRU)Ͱ •
GRUʹ͍࣋ͬͯͨ͘Ίɺ3D convolutionΛ༻ • ͜ͷաఔ͕MVSͷܭࢉϚονϯάΛ୲ • ֶशͷࡍը૾ͷೖྗॱΛϥϯμϜʹೖΕସ͑Δ 19
ख๏ ▸ 3D Grid Reasoning • GRUͰ̏࣍ݩάϦουʹͨ͠ΒϊΠζ͕ଟ͔ͬͨɻ • 3U-netͰEncode Decode͢ΔͱFilteringͰ͖Δ
20
ख๏ ▸ Differentiable Projection • Depthͷ෮ݩʹL1 loss(high frequency informationͷͨΊ) •
Voxelͷ෮ݩʹvoxel͝ͱͷcross entropy loss 21
࣮ݧ ▸ σʔληοτ • ShapeNetσʔλΛར༻ • ̏࣍ݩCADϞσϧͷެ։σʔληοτ 22 https://shapenet.cs.stanford.edu/shrec17/
࣮ݧ • ೖྗը૾ ▸ ShapeNetͷ3DϞσϧΛϨϯμϦϯάͯ͠224x224x3 ▸ ̍ࢹ͋ͨΓ̐ຕ ▸ Χϝϥϙʔζ •
Ξτϓοτ ▸ Depth: 224x224x3 ▸ Voxel: 32x32x32 23
࣮ݧ ▸ ݁Ռ 24 3DR2N2ͱൺɺࡉ͔͍෮ݩ͕Մೳ
࣮ݧ ▸ ݁Ռ 25 3DR2N2ͱൺɺগͳ͍ຕͰ෮ݩ͕Մೳ ຕ૿͑Δͱੑೳ্͕͕Δ
࣮ݧ ▸ ݁Ռ 26 stereo matchingͰ෮ݩ͠ͳ͍ ૭෮ݩՄೳ
࣮ݧ ▸ ݁Ռ 27 stereo matchingʹൺ গͳ͍ຕͰ෮ݩ͕Մೳ චऀᐌ͘ CNNͷίϯςΫετΛݟΔྗ ैདྷͷstereo
matchingΛ͙྇ DepthMapͷਪఆ݁ՌΛෳΈ߹Θͤͯ̏࣍ݩ෮ݩͨ͠
·ͱΊ ▸ Learnt Stereo MachinesΛఏҊ ▸ ෳࢹ͔Βͷೖྗը૾Λݩʹɺ DepthMapͱVoxelͷਪఆ͕Մೳͱͳͬͨ ▸ ՝
• ग़ྗVoxel͕32x32x32ͱখ͍͞ 28