Claudia Willimzig's wiki

Types of anomaly (Anomaly examples)

After we have a classification (we will probably end up with 8-10 classes) we could try and quantify the effect: how fast do subjects detect it, do subjects report it ... Unfortunately we will need more than 1 picture for each kind, otherwise it is difficult to generalize...

1. Low-level pop-out: (e.g. a fish in a school of fish is oriented differently)
2. object out of context (boat on road)
3. Physically impossible (boat floating in sky)
4. too much homogeneity (all the cars on the road are identical)
5. Too little homogeneity (e.g. a school of fish, but the fish all belong to different species)
6. Unlikely object (person with 1 leg only, mountain lion)
7. Another way to think about some of these anomalies: local (object is rare/impossible) vs global (object does not fit in scene)

Irv Biederman has a list of `inconsistencies' with nice names (grounding, ...). We should get a hold of the paper and read it. He has worked on this subject

After we have a classification (we will probably end up with 8-10 classes) we could try and quantify the effect: how fast do subjects detect it, do subjects report it ... Unfortunately we will need more than 1 picture for each kind, otherwise it is difficult to generalize...

Tasks for our subjects

1. Detect objects belonging to a given category (as in Wolfe's and Biederman's papers)
2. Detect object not belonging to given category
3. Detect anomaly or anomalous object
4. Free report (and we check whether anomaly is being reported)

Parameters of experiment

• Frequency of the anomaly
• Information from context (e.g. no context in Wolfe, context in Biederman)

Comparison to Wolfe's paper

Reading JM Wolfe's '05 Nature paper suggests a number of ways to think about your experiment:

a) his subjects were looking for something rare but specific (a tool), in your experiment you are asking the subjects to look for something rare and unexpected

b) in his experiment there was no `context': the objects were unrelated to one another; in your experiment there is a context and one of the tasks that your subjects may be carrying out is to detect something that is `out of context'.

So: one could ask these questions:

1. What happens if one repeats Wolfe's experiments, but one does not pre-specify the class of the odd objects to be detected (the non-odd objects could include e.g. household items). Would this make the task more difficult, or just the same?

2. Does context help in spotting odd objects? E.g. one could compare the task of spotting odd objects in context, with the task of spotting some class of objects: animals, tools, ...

3. What happens when the task is not specified (i.e. you do not tell the subjects that they are supposed to find odd objects) and all you ask is a report of what the subjects saw, after you presented the picture for, say, 0.5s.

It seems to me that there are at least thre different cues that your subjects could be using:

a) out-of-context objects: e.g. the sailing boat in the sky b) objects that are familiar and correctly placed in context, but are `damaged' or `odd' : e.g. the pedestrian with a leg missing, the rotated stop sign c) objects that are unfamiliar (suppose that an american subject saw a Trabant in Los Angeles)

Can you think of any other cues?

It would make sense to group your images by cue and study independently the effects.

Claudia's comments

I think, we should think about whether the picture itself might also establish a context in that sense that people roughly estimate what is "normal" in the context of the pic (normal size, normal width etc) (consistent with Treisman's work that people are very good at estimating averages) ... and then exceptions to this rule might be considered as anomalous events or at least get attention (f. e. anomalously large/small etc.). If indeed the definition of "odd" not only depends on long-term representations but on such within-scene characteristics this might lead to the contraintuitive hypothesis that people might actually be slower in rating something as anomalous if there are several anomalous events of the same kind (as a series of anomalies might establish their own "normality"). Or they would fail to detect anomalies if the anomalous events consists of abnormal similarities between objects (f. e. if they are all exactly the average -which would hardly ever occur in normal life). (please check the pics on pp. 3-8 in the attached file)

Re. the context: I also read another study by Hollingworth & Henderson (1998) (do you know it?) who actually put objects in line drawings into inconsistent context (e. g. a mixer in a farmyard). They find evidence that context does neither help nor perturb correct identification of objects.

I still think (Christof, we discussed this) that asking people to categorize objects might not be the best question; f. e. if we classify cars as cars based on the fact that cars have wheels and I take those wheels away - then it is hardly surprising that people take longer for classification. However, still their attention might be drawn to the car way earlier than that (as a "second glance" to confirm that it is still a car even without wheels). So I think at least for the start we might be better off with having them give spontaneous reports or measuring how long it takes them to detect an anomalous event.

Literature

1. Biederman's papers on object detection in consistent / inconsistent surroundings
2. Wolfe's Nature 2005 paper
3. Fei Fei Li's JOV paper on gist.
4. ...