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
[RSJ22] TDP-MAT: Multimodal Language Comprehens...
Search
Sponsored
·
Ship Features Fearlessly
Turn features on and off without deploys. Used by thousands of Ruby developers.
→
Semantic Machine Intelligence Lab., Keio Univ.
PRO
September 05, 2022
Technology
840
0
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
[RSJ22] TDP-MAT: Multimodal Language Comprehension for Object Manipulation Tasks via Realย Images
Semantic Machine Intelligence Lab., Keio Univ.
PRO
September 05, 2022
More Decks by Semantic Machine Intelligence Lab., Keio Univ.
See All by Semantic Machine Intelligence Lab., Keio Univ.
[Journal club] Predict Before You Explore: Predictive Planning with Specialized Memory for Embodied Question Answering
keio_smilab
PRO
0
81
[Journal club] PHyCLIP: ๐๐-Product of Hyperbolic Factors Unifies Hierarchy and Compositionality in Vision-Language Representation Learning
keio_smilab
PRO
0
86
[Journal club] ReMEmbR: Building and Reasoning Over Long-Horizon Spatio-Temporal Memory for Robot Navigation
keio_smilab
PRO
0
110
[Journal club] ReLaGS: Relational Language Gaussian Splatting
keio_smilab
PRO
0
120
[Journal club] Flow as the Cross-Domain Manipulation Interface
keio_smilab
PRO
0
98
Mobi-๐: Mobilizing Your Robot Learning Policy
keio_smilab
PRO
0
160
A Gentle Introduction to Transformers
keio_smilab
PRO
16
7k
FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
keio_smilab
PRO
0
60
[Journal club] VLA-Adapter: An Effective Paradigm for Tiny-Scale Vision-Language-Action Model
keio_smilab
PRO
1
150
Other Decks in Technology
See All in Technology
็ฌฌ67ๅใณใณใใฅใผใฟใใธใงใณๅๅผทไผCVPR2026่ชญไผๅ็ทจ
tsukamotokenji
0
160
ๅคๆดใ็ถใใใใใทในใใ ใใฉใไฟใคใ โ AIๆไปฃใฎSSoTใจใใ่จญ่จๅๅ
kawauso
1
1.1k
CDKใงๆธใECSใฎใในใใใฉใฏใใฃในใ ๆนใใฆ่ใ็ดใ2026 #cdkconf2026
makies
3
940
Multicaใง30ๅใฎใใใใญใธใงใฏใใAIใจใผใธใงใณใ้็จใใฆ่ฆใใฆใใใใจ
eiei114
1
630
AIๆไปฃใซใใใใในใใฎๅบ็คใฎๅๅฎ็พฉ / Rethinking the Fundamentals of Testing in the AI Era
mineo_matsuya
13
4.4k
ใใผใฟใจๅฐๅณใง่ชญใ ๅคงไบ็บใฎใใใใใใฎใใใใใชใใใฎใ
yoshiyama_hana
0
140
AIๆไปฃใใใในใฑใผใซใใชใใใจใใใใ -ใไฝใไบบใใใใใชใไฝใใใใ่ใใไบบใธ / Do Things That Don't Scale in the AI Era โ From How to Why
kaminashi
1
100
ใใขใใซ + ใใผใในใใง่ชญใฟ่งฃใ AIใจใผใธใงใณใๅ ฅ้
oracle4engineer
PRO
2
160
ใใผใฟใจใณใธใใขใชใณใฐใจใใกใคใณ้งๅ่จญ่จ
masuda220
PRO
14
2.5k
่คๆฐใใญใใฏใใง้ฒใใAIๆฉ่ฝๅฎ่ฃ โโ ๅฎ่ทตใใๅพใใชใขใซใชๅญฆใณใจใญใผใใใใๅฎ็พใธใฎๆๆฆ / AICon2026_yanari
rakus_dev
1
270
ใชใใใใชใใฎAPIใฏไฝฟใใใชใใฎใ๏ผ AXๆไปฃใฎ่จญ่จๅๅใใฌใผใใฌใผใซใ้็จไฝๅถ
yokawasa
1
210
Vใใคใณใๅๆๅบ็คใซใใใใใผใฟใขใใชใณใฐ20ๅนดๅฒ
taromatsui_cccmkhd
4
720
Featured
See All Featured
AI: The stuff that nobody shows you
jnunemaker
PRO
9
840
The Invisible Side of Design
smashingmag
301
52k
A Soul's Torment
seathinner
6
3.1k
Marketing to machines
jonoalderson
1
5.6k
Faster Mobile Websites
deanohume
310
32k
Building Better People: How to give real-time feedback that sticks.
wjessup
370
20k
Crafting Experiences
bethany
1
230
Between Models and Reality
mayunak
4
380
The browser strikes back
jonoalderson
0
1.4k
Refactoring Trust on Your Teams (GOTO; Chicago 2020)
rmw
35
3.7k
JAMstack: Web Apps at Ludicrous Speed - All Things Open 2022
reverentgeek
1
500
Improving Core Web Vitals using Speculation Rules API
sergeychernyshev
21
1.5k
Transcript
1
2
3 โ https://www.toyota.com/usa/toyota-effect/romy-robot
โ : 4 โLook in the left wicker vase that
is next to the potted plantโ Wicker vase :
โ : โLook in the left wicker vase that is
next to the potted plantโ 5 Wicker vase : Wicker vase Wicker vase Wicker vase
โ : โ Key : โLook in the left wicker
vase that is next to the potted plantโ 6 Wicker vase : Wicker vase Wicker vase Wicker vase
โ REVERIE-fetch โข 7 โLook in the left wicker vase
that is next to the potted plantโ
โ REVERIE-fetch โข โข (Instruction) (Context Regions) (Candidate Region) 8
โLook in the left wicker vase that is next to the potted plantโ
โ REVERIE-fetch โข โข (Instruction) (Context Regions) (Candidate Region) 9
โLook in the left wicker vase that is next to the potted plantโ
โ REVERIE-fetch โข โข (Instruction) (Context Regions) (Candidate Region) โข
10 โLook in the left wicker vase that is next to the potted plantโ
โ REVERIE-fetch โข โข (Instruction) (Context Regions) (Candidate Region) โข
11 โLook in the left wicker vase that is next to the potted plantโ Faster R-CNN[Ren+, PAMI16]
MTCM [Magassouba+, RA-L19] . VGG16LSTM . Target-dependent UNITER (TDU) [Ishikawa+,
RA-L21] UNITER[Chen+, ECCV20] . REVERIE task / dataset [Qi+, CVPR20] , REVERIE 12
โข MAT[Ishikawa+, ICPR22] โข CLIP[Radford+, ICML21] โข Perceiver[Jaegle+, ICML21] 13
โข MAT[Ishikawa+, ICPR22] โข CLIP[Radford+, ICML21] โข Perceiver[Jaegle+, ICML21] 14
โข MAT[Ishikawa+, ICPR22] โข CLIP[Radford+, ICML21] โข Perceiver[Jaegle+, ICML21] 15
โข MAT[Ishikawa+, ICPR22] โข CLIP[Radford+, ICML21] โข Perceiver[Jaegle+, ICML21] 16
โข MAT[Ishikawa+, ICPR22] โข CLIP[Radford+, ICML21] โข Perceiver[Jaegle+, ICML21] 17
2 1 3
โ ๐น๐ก โ 18 Input ๐น๐ก Output 1. ๐ธ ๐น
= CE ๐ ๐ , ๐ โ๐น ๐ธ ๐น = ๐๐ธ ๐๐น 2. โ๐น ๐ธ ๐น ๐๐ก ๐๐ก ๐๐ก = ๐1 ๐๐กโ1 + 1 โ ๐1 โ๐น ๐ธ ๐น๐ก ๐๐ก = ๐2 ๐๐กโ1 + 1 โ ๐2 โ๐น ๐ธ ๐น๐ก 2 3. ๐๐ก ๐๐ก โ๐น๐ เท ๐๐ก = ๐๐ก 1 โ ๐1 ๐ก , เท ๐๐ก = ๐๐ก 1 โ ๐2 ๐ก โ๐น๐ = ๐ เท ๐๐ก เท ๐๐ก + ๐ 4. ๐น๐ก+1 = ฮ ๐น โค๐ ๐น๐ก + โ๐น๐ โ๐น๐ ๐น
โ CLIP โ ViT[Dosovitskiy+, ICLR21] โ transformer [EOT] 19 [EOT]
โ โ Perceiver CLIP 20 CLIP Encoders
โ CLIP Encoders , Perceiver 21
โ REVERIE-fetch dataset - REVERIE dataset โ REVERIE[Qi+, CVPR18] -
โ 1. , 2. https://yuankaiqi.github.io/REVERIE_Challenge/static/img/demo.gif 22 Matterport3D
โ REVERIE-fetch dataset - REVERIE dataset โ REVERIE[Qi+, CVPR18] :
+ 23 , โ - REVERIE - - https://yuankaiqi.github.io/REVERIE_Challenge/static/img/demo.gif
โ REVERIE-fetch dataset โข REVERIE dataset #Samples Vocabulary size Average
sentence length 30532 2853 19.1 Training Validation Test 26808 2552 1172 24 โLook in the left wicker vase that is next to the potted plantโ
โGo into the living room and give me the pillow
on the couch nearest the plantโ 25 โข โ TDP-MAT
26 โข โ TDP-MAT โ Bounding box โMake haste to
the office and fluff the pillow sitting on the left of the chairโ
โข Acc [%] : 27 Condition Acc [%] โ Baseline
: TDU [Ishikawa+, IROS21] 73.3 0.485 Ours : TDP-MAT W/o MAT 72.5 3.55 W/o MAT + Smaller learning rate 74.4 0.831 W/o CLIP & Perceiver 74.1 1.47 W/o Pretraining 73.1 2.24 Full 75.3 0.691 +2.0
28 Condition Acc [%] โ Baseline : TDU [Ishikawa+, IROS21]
73.3 0.485 Ours : TDP-MAT W/o MAT 72.5 3.55 W/o MAT + Smaller learning rate 74.4 0.831 W/o CLIP & Perceiver 74.1 1.47 W/o Pretraining 73.1 2.24 Full 75.3 0.691 +2.8 - - 5 - ( ) - Smaller learning rate : 1/8 -
29 Condition Acc [%] โ Baseline : TDU [Ishikawa+, IROS21]
73.3 0.485 Ours : TDP-MAT W/o MAT 72.5 3.55 W/o MAT + Smaller learning rate 74.4 0.831 W/o CLIP & Perceiver 74.1 1.47 W/o Pretraining 73.1 2.24 Full 75.3 0.691 +1.2 - CLIP Encoders, Perceiver Module, - Cross Attention
30 Condition Acc [%] โ Baseline : TDU [Ishikawa+, IROS21]
73.3 0.485 Ours : TDP-MAT W/o MAT 72.5 3.55 W/o MAT + Smaller learning rate 74.4 0.831 W/o CLIP & Perceiver 74.1 1.47 W/o Pretraining 73.1 2.24 Full 75.3 0.691 +2.2 - TDU
โ โข โ โข MAT โข โ โข 31
โ โ ๐ฟ ๐ ๐ ๐ฟร๐ท ๐ ๐ร๐ธ ๐ ๐ฟร๐ท, ๐ ๐ร๐ท โ ๐ ๐ฟร๐
๐ ๐ฟร๐ท ๐ ๐ฟร๐ท, ๐ ๐ฟร๐ท โ ๐ ๐ฟร๐ฟ 32
โ โ โ โ 33
โ 34 8 ร 10โ4 ๐ฝ1 = 0.9, ๐ฝ2 =
0.99
โ โ โ 35 19+6=25