Second Interna,onal Workshop on Parts and A5ributes ECCV 2012, Firenze, Italy October, 2012 Discovering a Lexicon of Parts and Attributes

Size: px

Start display at page:

Download "Second Interna,onal Workshop on Parts and A5ributes ECCV 2012, Firenze, Italy October, 2012 Discovering a Lexicon of Parts and Attributes"

Nora Preston
6 years ago
Views:

1 Second Interna,onal Workshop on Parts and A5ributes ECCV 2012, Firenze, Italy October, 2012 Discovering a Lexicon of Parts and Attributes Subhransu Maji Research Assistant Professor Toyota Technological Institute at Chicago

2 Motivation Detailed object recognition Communication requires a lexicon Diverse Visual Categories CUB 200 dataset,visipedia project High%heel(Blue(Shoe( Berg et al., 10 Farhadi et al, 09

3 Source of part and attribute lexicons Field guides provide exhaustive lists when available Expert vs. Layman Task specific vs. not

4 Source of part and attribute lexicons Captioned images Limited by sources of such text Descriptions are often not visual Image from Berg et al., ECCV 10

5 What are good attribute lexicons? Properties It should be easy to communicate It should be easy to differentiate instances from one another images from google, cub 200

6 Discriminative description task Describe the (visual) differences between the two

7 Discriminative description task

8 Discriminative description task Description!"#$%&'(&)'*)#% %&'()$ *'"$)(+,()$ -).$/0&0-$ *'"$-1..)-$

9 Discriminative description task Description Discriminative Description!"#$!"#$%&'(&)'*)#% %&'()$ *'"$)(+,()$ -).$/0&0-$ *'"$-1..)-$!"#$%+",)')-.)#% $%-0%)&&)-$%&'()$!"#$%'"")(+)-$%&'()$ $$0()$)(+,()$!"#$201-$)(+,()"$ $$$$-).$/0&0-$!"#$3*,4)$/0&0-$ -01(.$-1..)-$!"#$%0,(45$-1..)-$ Helps elicit a lexicon that enables fine grained discrimination Is task specific by design

10 Discriminative description task Description Discriminative Description!"#$!"#$%&'(&)'*)#% %&'()$ *'"$)(+,()$ -).$/0&0-$ *'"$-1..)-$!"#$%+",)')-.)#% $%-0%)&&)-$%&'()$!"#$%'"")(+)-$%&'()$ $$0()$)(+,()$!"#$201-$)(+,()"$ $$$$-).$/0&0-$!"#$3*,4)$/0&0-$ -01(.$-1..)-$!"#$%0,(45$-1..)-$ Helps elicit a lexicon that enables fine grained discrimination Is task specific by design Non-accidental properties, Biederman 87 Pragmatics of Language, Levinson 83

11 Collecting descriptions on MTurk Interface on AMT Example annotations free form text (separated by vs ) Minimizes instruction bias Keeps the interface simple

hold door open" propeller to the body one rudder thin body low wings facing towards left side

12 Example annotations: airplanes facing left turbofan powered plane longer tail green rudder passenger door open facing right propeller powered plane shorter tail white rudder baggage hold door open" propeller to the body one rudder thin body low wings facing towards left side propeller to the wing two rudders fat body high wings facing slightly towards" Images from airliners.net

black body pointy beak short tail black spot over head short leg orange brown body

13 Example annotations: birds black and white wings white body large eyes small tail v shaped beak spotted wings spotted body small eyes long tail pointed beak" yellow black body pointy beak short tail black spot over head short leg orange brown body shape beak long tail brown stripe over head long leg" Images from CUB 200 dataset

vs. flying feather color tail size leg color

14 Analyzing the text: instance specific properties bird family fur color beak shape tail size beak color beak size tail size wing color head color leg color beak color sitting vs. flying feather color tail size leg color Different properties are revealed for different instances

15 Analyzing the text: instance specific properties Frequency of usage is a measure of its discriminability

16 Analyzing the text: instance specific properties Frequency of usage is a measure of its discriminability

17 Analyzing the text : discovering a lexicon red rudder vs. white rudder pointy nose vs. round nose sentence pairs!"#$!"#$%+",)')-.)#% $%-0%)&&)-$%&'()$!"#$%'"")(+)-$%&'()$

18 Analyzing the text : discovering a lexicon!"#$!"#$%+",)')-.)#% $%-0%)&&)-$%&'()$!"#$%'"")(+)-$%&'()$ red rudder vs. white rudder pointy nose vs. round nose sentence pairs rudder nose parts {red, white} {pointy, round} modifiers

19 Analyzing the text : discovering a lexicon!"#$!"#$%+",)')-.)#% $%-0%)&&)-$%&'()$!"#$%'"")(+)-$%&'()$ red rudder vs. white rudder pointy nose vs. round nose sentence pairs rudder nose parts {red, white} {pointy, round} modifiers Key idea nouns : words that repeat modifiers : words that are different Each sentence has only one noun and modifier

20 Analyzing the text : discovering a lexicon Sentence alignment yellow beak black and white beak Used in NLP to initialize translation tables (IBM models)

21 Analyzing the text : discovering a lexicon Sentence alignment yellow beak black and white beak Used in NLP to initialize translation tables (IBM models) Goals Discover a lexicon of parts, modifiers and part-modifier relations Modifiers should be shared across attributes Estimate the frequency of each attribute

22 A generative model of sentence pairs

23 A generative model of sentence pairs J a f z I t e N

24 A generative model of sentence pairs J a f yellow beak z I t e vs. black and white beak N

25 A generative model of sentence pairs J a f yellow beak z I t e vs. black and white beak N part-modifier topic pairs beak-size beak-color bird-size wing-color bird-kind leg-color head-color...

26 A generative model of sentence pairs J a f yellow beak z I t e vs. black and white beak N part-modifier topic pairs beak-size beak-color bird-size wing-color bird-kind leg-color head-color... z beak-color

27 A generative model of sentence pairs J a f yellow beak z I t e vs. black and white beak N part-modifier topic pairs beak-size beak-color bird-size wing-color bird-kind leg-color head-color... z beak-color t topic:color topic:beak

28 A generative model of sentence pairs J a f yellow beak z I t e vs. black and white beak N part-modifier topic pairs beak-size beak-color bird-size wing-color bird-kind leg-color head-color... z beak-color t topic:color topic:beak e yellow beak NULL

29 A generative model of sentence pairs J a f yellow beak z I t e vs. black and white beak N part-modifier topic pairs beak-size beak-color bird-size wing-color bird-kind leg-color head-color... z beak-color t topic:color topic:beak e yellow beak NULL a

30 A generative model of sentence pairs J a f yellow beak z I t e vs. black and white beak N part-modifier topic pairs beak-size beak-color bird-size wing-color bird-kind leg-color head-color... z beak-color t topic:color topic:beak e yellow beak NULL a f black and white beak

31 A generative model of sentence pairs J a f yellow beak z I t e vs. black and white beak N part-modifier topic pairs beak-size beak-color bird-size wing-color bird-kind leg-color head-color... z t e a f beak-color topic:color topic:beak yellow beak NULL black and white beak Initialize part and modifier topics using word alignments

32 Parts, modifiers and attributes of airplanes wheel wheels plane engine engines rudder wings wing front back nose facing body tail GLOBAL one two no single three double four color black sky light whiteblue ordinary colored dark whitegreen whitered pointy round flat pointed sharp point square propeller passenger jet only military cargo small big large medium white red blue green yellow gray orange brown top bottom middle down open closed opened close right left slightly on near off 200 images, 1000 random pairs Images from airliners.net

33 Parts, modifiers and attributes of airplanes 1 wheel wheels plane engine engines rudder wings wing front back nose facing body tail GLOBAL one two no single three double four color black sky light whiteblue ordinary colored dark whitegreen whitered pointy round flat pointed sharp point square propeller passenger jet only military cargo small big large medium white red blue green yellow gray orange brown top bottom middle down open closed opened close right left slightly on near off 200 images, 1000 random pairs Images from airliners.net

34 Parts, modifiers and attributes of airplanes 2 1 wheel wheels plane engine engines rudder wings wing front back nose facing body tail GLOBAL one two no single three double four color black sky light whiteblue ordinary colored dark whitegreen whitered pointy round flat pointed sharp point square propeller passenger jet only military cargo small big large medium white red blue green yellow gray orange brown top bottom middle down open closed opened close right left slightly on near off 200 images, 1000 random pairs Images from airliners.net

35 Parts, modifiers and attributes of airplanes wheel wheels plane engine engines rudder wings wing front back nose facing body tail GLOBAL one two no single three double four color black sky light whiteblue ordinary colored dark whitegreen whitered pointy round flat pointed sharp point square propeller passenger jet only military cargo small big large medium white red blue green yellow gray orange brown top bottom middle down open closed opened close right left slightly on near off 200 images, 1000 random pairs Images from airliners.net

36 Parts, modifiers and attributes of airplanes localized vs. global wheel wheels plane engine engines rudder wings wing front back nose facing body tail GLOBAL one two no single three double four color black sky light whiteblue ordinary colored dark whitegreen whitered pointy round flat pointed sharp point square propeller passenger jet only military cargo small big large medium white red blue green yellow gray orange brown top bottom middle down open closed opened close right left slightly on near off 200 images, 1000 random pairs Images from airliners.net

37 Parts, modifiers and attributes of birds beak-size, wing-color, tail-size, body color, bird type bird wings feather feathers tail beak like body leg legs eyes neck head in fur GLOBAL long short small large big v sparrow pointy duck pointed sparow point crow bend eagle bended dove pointly pigeon brown blue yellow gray red green spotted slightlyhumming ash kite light parrot black white orange fat slim silm lean sharp round flat normal shaped curved blunt little rounded ordinary 200 images, 1600 pairs 1 image per category CUB 200 dataset

38 Parts, modifiers and attributes of people location, shirt vs tshirt, hair length, gender hand hands hair facing snap snape the picture spectacles in glass bag watch glasses GLOBAL towards left right backwards forward sidewards man lady woman boy girl ladies baby adult child kid children adults shirt tshirt jacket dress coat tshirts wearing not smiling having asian side caucasian back africans front asians backside latin frontal western toward turn turned rear sideways upright us black white blue brown blonde red green gray yellow colored pink orange single couple 2 non double group couples many 3 dark light fair tee sky show bright lighter design medium thick fat slim normal thin lean fit average skinny female male full half of sleeve only fully sleeveless jeens close partial rain thinning torso waist indoor outdoor door homely stage inside outside out handed long short small shot women men womens young old older middle mature elderly two sitting one standing three walking both riding somebody cycling weight sleeping dancing driving no with without alone bald 400 images, 1600 pairs Random images from PASCAL VOC 11

39 Conclusions Discriminative description is an effective way to elicit a lexicon of attributes that are useful for fine-grained distinction Simple analysis of sentence pairs can help discover a lexicon of parts a lexicon of modifiers the relative frequency of these in a dataset relationships between parts and modifiers (attributes) A useful tool to bootstrap annotation collection process

C R H G E K. 1 Solve the puzzle. lion. parrot. crocodile. flamingo. snake. tortoise. horse. zebra. elephant. eagle duck. monkey. Classify the animals.

C R H G E K. 1 Solve the puzzle. lion. parrot. crocodile. flamingo. snake. tortoise. horse. zebra. elephant. eagle duck. monkey. Classify the animals. Wild animals 1 Solve the puzzle. Z E C R lion crocodile snake parrot H G 0 E L 0 E K flamingo tortoise D horse zebra elephant Classify the animals. monkey eagle duck Mammals Birds Reptiles 40 Unit 4 Wild