Idanwo ti o sopọ mọ itọsọna · Alabọde Ipele

Ibeere Wiwo Idahun Idahun

Review visual question answering, dataset bias, multimodal fusion and evaluation limits across image and document questions.

Jẹmọ itọsọna awọn ọnaIdahun Ibeere wiwo
Ibeere 1 ti 8

Awọn igbewọle meji wo ni eto Idahun Ibeere Wiwo kan gba?

Awotẹlẹ gbogbo ibeere8 ibeere

Awọn idahun ati awọn alaye han bi o ṣe n mu ibeere naa.

  1. Awọn igbewọle meji wo ni eto Idahun Ibeere Wiwo kan gba?
    • Aworan ati ibeere ede adayeba
    • Awọn aworan meji lati ṣe afiwe
    • Ibeere ati agekuru ohun
    • Aworan nikan ti ko si ọrọ
  2. Kini idi ti dataset VQA v2.0 ti ṣẹda pẹlu awọn idahun 'iwọntunwọnsi'?
    • Lati ṣe awọn aworan ti o ga
    • Lati dinku abosi ede nitoribẹẹ awọn awoṣe ko le gboju lati ọrọ nikan
    • Lati yọ gbogbo bẹẹni/ko si ibeere
    • Lati dinku dataset fun ikẹkọ yiyara
  3. Kini 'aiṣedeede ede' ni VQA?
    • Itumọ awọn ibeere si ede ti ko tọ
    • Lilo awọn ede pupọ ni ẹẹkan
    • Idahun ti o da lori ọrọ ibeere lakoko ti o kọju si aworan naa
    • Yiyan awọn aworan kukuru ju awọn ti o gun lọ
  4. Bawo ni ọpọlọpọ awọn awoṣe VQA ṣe darapọ aworan ati alaye ibeere?
    • Nipa pipọ awọn piksẹli ati awọn lẹta aise
    • Nipa ṣiṣẹda aworan tuntun lati ibeere naa
    • Nipa aibikita ibeere naa patapata
    • Nipa didi awọn aṣoju ti a fi koodu si wọn, nigbagbogbo pẹlu akiyesi agbelebu
  5. Awọn ọna ṣiṣe VQA Ayebaye nigbagbogbo ṣe awọn idahun nipasẹ ṣiṣe kini?
    • Pinpin lori awọn fokabulari ti o wa titi ti awọn idahun ti o wọpọ
    • Yiya apoti didi nikan
    • Nsatunkọ awọn atilẹba aworan
    • Nda ibeere titẹ sii pada ko yipada
  6. Ewo ni apẹẹrẹ ti ipenija VQA ti ilọsiwaju diẹ sii?
    • Yiyipada aworan si eekanna atanpako
    • Idahun awọn ibeere nipa ọrọ ati awọn nọmba inu aworan iwe-ipamọ kan
    • Tito awọn aworan lẹsẹsẹ nipasẹ orukọ faili
    • Fisinuirindigbindigbin awọn faili fidio
  7. How does the original VQA benchmark score an answer when people phrase the same answer differently?
    • It requires an exact match to one fixed phrase only
    • It scores image resolution instead of answers
    • It compares the prediction with multiple human answers using a consensus-based score
    • It uses the model’s confidence as the ground truth
  8. Why can high VQA accuracy still fail to prove visual grounding?
    • Images cannot be represented by neural networks
    • Questions are always randomly generated
    • The benchmark contains no answer labels
    • A model may exploit dataset or language regularities instead of the image