Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner &...

16
Shadows (P) Lavanya Sharan February 21st, 2011
  • date post

    20-Dec-2015
  • Category

    Documents

  • view

    218
  • download

    1

Transcript of Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner &...

Page 1: Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner & Ramchandran 1992Enns & Rensink 1990.

Shadows (P)Lavanya Sharan

February 21st, 2011

Page 2: Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner & Ramchandran 1992Enns & Rensink 1990.

Anomalously lit objects are easy to see

Kleffner & Ramchandran 1992

Enns & Rensink 1990

Page 3: Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner & Ramchandran 1992Enns & Rensink 1990.

Anomalously lit objects are easy to see

Kleffner & Ramchandran 1992

Enns & Rensink 1990

Page 4: Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner & Ramchandran 1992Enns & Rensink 1990.

Anomalously lit objects are easy to see

Ostrovsky et al. (2005): Only in these conditions.

Kleffner & Ramchandran 1992

Enns & Rensink 1990

Page 5: Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner & Ramchandran 1992Enns & Rensink 1990.

Notion of distractor homogeneity

Enns & Rensink (1990)

Target is lit from belowDistractors are lit from above

Task-relevant dimension = Light direction

Here, distractors are identical to each other => Complete distractor homogeneity.

Search tasks are easier when distractors are homogeneous.

Page 6: Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner & Ramchandran 1992Enns & Rensink 1990.

Notion of distractor homogeneity

Ostrovsky et al. (2005)

Target is lit from the sideDistractors are lit from above

Task-relevant dimension = Light direction

Here, distractors differ from each other in orientation => Decreased distractor homogeneity.

Task-irrelevant dimension = Orientation

Search tasks can get harder when distractors are

inhomogeneous.

Page 7: Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner & Ramchandran 1992Enns & Rensink 1990.

Ostrovsky et al. experiment

Ostrovsky et al. (2005)

Target and distractor lighting differ by 90 degrees.

Lighting conditions vary in 45 deg steps from 0 to 360.

Distractor cubes vary in orientation.

Number of items = 4, 9 and 12.

Presented for 100, 500 and 1000 ms.

Page 8: Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner & Ramchandran 1992Enns & Rensink 1990.

Search for anomalous illumination is not efficient

Ostrovsky et al. results contradict previous work.

When reaction time increases and accuracy decreases with number of items = signature of serial search (opposite of ‘pop-out’).

Page 9: Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner & Ramchandran 1992Enns & Rensink 1990.

What is going on?

No. They ran a baseline condition where distractors were identical and reproduced ‘pop-out’-like performance.

90% accuracy at 120 ms display time (chance = 50%).Performance invariant to number of items (4-12).

Something weird about their setup?

Page 10: Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner & Ramchandran 1992Enns & Rensink 1990.

What is going on?

Of course, distractor inhomogeneity makes a task hard. This is a stress test.

Perhaps. But,

i) Real world has a lot of inhomogeneity.ii) They conducted an experiment where task-relevant dimension was cube shape, and task-irrelevant dimension (as before) was orientation.

At 1000 ms, number of items = 9Shape task 92%, Illumination task 56% (chance = 50%)

Page 11: Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner & Ramchandran 1992Enns & Rensink 1990.

What is going on?

Advantage for top and top-left conditions, but even in those no pop-out.

Wait, we know there is a bias for top-left. They tested all directions, perhaps pop-out is in the top-left conditions.

Page 12: Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner & Ramchandran 1992Enns & Rensink 1990.

What is going on?

Cube shapes are weird. Ran experiment with other shapes, same results. Ran a baseline to confirm participants can estimate illumination direction from these new shapes.

Page 13: Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner & Ramchandran 1992Enns & Rensink 1990.

Second study with real-world stimuliDigitally modified images to have inconsistent illlumination (average diff = 90 deg)

This is clearly not pop-out.

When not explicitly instructed to look for illumination inconsistency, performance was at chance.

Page 14: Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner & Ramchandran 1992Enns & Rensink 1990.

Support from other studies

When stimuli can be seen as object + shadow vs. object + a second attached object, search is slower in shadow case.

Rensink & Cavanagh (2004)

Page 15: Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner & Ramchandran 1992Enns & Rensink 1990.

Support from other studies

Farid & Bravo (2010)When cast shadows in opposite directions, near perfect

detection.When cast shadows in same direction, near chance

performance.

Page 16: Shadows (P) Lavanya Sharan February 21st, 2011. Anomalously lit objects are easy to see Kleffner & Ramchandran 1992Enns & Rensink 1990.

Conclusions

✓ Global illumination errors (and therefore shadow consistencies) are hard to detect, especially in complex scenes.

✓ If we are estimating illumination really well (and therefore estimating shape & reflectance really well), why do we make these mistakes? Inverse optics theories have to account for these errors.

✓ Unclear whether these mistakes happen because we are bad at estimating illumination (imperfect inverse optics) vs. unable to report correctly (read out issue).