Scribd
Upload
3K views4 pages

ATLAS Quiz: Labeling Guidelines Explained

The document provides guidelines for labeling actions in a knowledge quiz context, emphasizing the importance of accuracy and clarity in descriptions. It covers various scenarios, including how to label actions, the use of timestamps, and the acceptable formats for labels. Key points include capturing both actions in labels, avoiding the combination of 'No Action' with real actions, and ensuring consistent terminology for objects.

Uploaded by

bonfacemutie39
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
3K views4 pages

ATLAS Quiz: Labeling Guidelines Explained

The document provides guidelines for labeling actions in a knowledge quiz context, emphasizing the importance of accuracy and clarity in descriptions. It covers various scenarios, including how to label actions, the use of timestamps, and the acceptable formats for labels. Key points include capturing both actions in labels, avoiding the combination of 'No Action' with real actions, and ensuring consistent terminology for objects.

Uploaded by

bonfacemutie39
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd

ATLAS KNOWLEDGE QUIZ ANSWERS

1. How should labels be written?  "Pick up mat and place on table" or


"Move mat to table" – must capture
 As passive observations (e.g., "the
both actions ✅
cup was picked up")
 Any of the above are acceptable
 As direct instructions (e.g., "pick up
the cup") ✅ Correct answer: Capture both actions

 As timestamps only

 As questions (e.g., "is the cup being 4. What is required for timestamps to be
picked up?") correct?

Correct answer: As direct instructions  They can be off by a few seconds

 Begin when hands reach; end when


hands disengage; must be accurate
2. The subject washes their hands, checks
✅
their watch, and adjusts their head-
mounted camera. How should this be  Only the start timestamp needs to
labeled? be accurate

 "wash hands, check watch, adjust  Timestamps are optional


camera"
Correct answer: Accurate start and end
 "wash hands" – the most visible based on hand movement
action

 "no action" ✅
5. Is it acceptable to use sequence words
 "prepare for task" like "another" and "next"?

Correct answer: "no action"  No – each action must be


independent

 Yes – if each action and timestamp is


3. The ego picks up a mat from the floor
accurate ✅
and places it on the table. Which label is
acceptable?  Only for actions repeated more than
3 times
 "Place mat on table" – the goal is
captured  Only at the beginning of a video

 "Pick up mat" – the first action is Correct answer: Yes, if accurate


captured
ATLAS KNOWLEDGE QUIZ ANSWERS

6. When multiple events happen at once, 9. Can you combine "No Action" with other
what must the labels capture? actions in one label?

 Only the first action  Yes – "No Action, pick up cup"

 Only the most important action  No – never combine No Action with


real actions ✅
 All actions with meaningful object
interactions ✅  Only if No Action comes first

 A one-word summary  Only for long segments

Correct answer: All meaningful object Correct answer: Never combine


interactions

10. The ego picks up items while closing a


7. The subject picks up a red mug from the drawer with their hip. What should be
sink while a blue mug sits on the counter. labeled?
Which label is correct?
 "Close drawer, pick up items"
 "pick up mug"
 "Pick up items" only ✅
 "pick up mug from sink" or "pick up
 "No Action"
red mug" ✅
 "Close drawer" only
 "grab item from sink"
Correct answer: Hand-based actions only
 "pick up blue mug"

Correct answer: Clearly identify the correct


object 11. The subject adjusts the corners of a t-
shirt to see the design. Which label is
appropriate?
8. What is the maximum word count for a
 "inspect t-shirt"
single dense label segment?
 "adjust t-shirt" or "straighten t-shirt"
 10 words
✅
 15 words
 "look at t-shirt"
 20 words (about 4 atomic actions) ✅
 "examine t-shirt design"
 No limit
Correct answer: Physical manipulation
Correct answer: 20 words
ATLAS KNOWLEDGE QUIZ ANSWERS

12. Are "lift", "pick up", and "grab"  No – must specify hand
interchangeable?
 No – must describe book
 No – use consistent terminology ✅
 Yes – valid annotation ✅
 Yes – flexibility is encouraged
 Only for short videos
 Only if ambiguous
Correct answer: Yes
 Only for specific objects

Correct answer: No
16. Can relative positions like "left" and
"right" be used?

13. When should a label be "idle" or "no  No – never


action"?
 Yes – if accurate from ego’s POV ✅
 When ego is walking or hands are
 Only for stationary objects
not interacting with objects ✅
 Only with duplicates
 When ego speaks
Correct answer: Yes, if accurate
 When ego looks away

 Always at start/end
17. The ego appears about to put down a
Correct answer: No hand–object interaction
phone but doesn’t. Can you label it as "put
down phone"?

14. What is the primary focus of  Yes – intent is clear


annotations?
 No – only label observable actions ✅
 Ego movement
 Yes – if phone is visible
 Main actions and hand dexterity
 Only if within 2 seconds
relevant to the task ✅
Correct answer: Only label what happens
 Every action

 Only electronics

Correct answer: Task-relevant hand actions


18. The ego lifts a shirt and shakes it out.
Which labels are acceptable?
15. Is "pick up book" acceptable without
 Only exact wording
specifying the hand?
ATLAS KNOWLEDGE QUIZ ANSWERS

 Multiple accurate descriptions are


acceptable ✅

 "Lift shirt" only

 Must include hand positions

Correct answer: Multiple accurate


descriptions

19. A label says "pick up weight and frame"


but only the weight was picked up. What
should happen?

 Accept

 Edit to "pick up weight" ✅

 Add "attempted frame"

 Accept if timestamps match

Correct answer: Edit to match reality

20. If one annotator says "orange box" and


another says "yellow box", is this a
problem?

 Yes – must be correct color ✅

 No – inconsistency is fine

 Yes – same terminology required

 Depends on video quality

Correct answer: Yes, orange and yellow are


different

Common questions

Powered by AI

A label is considered 'jointly capturing' multiple actions if it explicitly documents both or all actions involved, such as 'Pick up mat and place on table' or 'Move mat to table'. This captures the entire sequence of actions involving the object, providing a comprehensive description of the interaction .

The primary consideration when labeling hand actions is to ensure that the labels capture the main actions and hand dexterity relevant to the task. It focuses on the interaction with objects and the task at hand rather than general movements or environmental descriptions .

The use of relative positions like 'left' and 'right' is permissible if they are accurate from the ego's point of view. This ensures that the labels are contextually relevant and correctly oriented to the perspective of the observer .

An action should be labeled as 'idle' or 'no action' when the ego is walking or when their hands are not interacting with objects. These labels are indicative of circumstances where there is a lack of hand-object interaction .

Yes, the use of sequence words like 'another' and 'next' is permissible in annotations as long as each action and timestamp is accurate. It ensures clarity in the sequence of events and facilitates a structured understanding of the actions .

An action should not be annotated if the ego is about to perform it but ultimately does not; only observable actions should be labeled. This rule avoids misleading annotations based on intent or assumptions about unexecuted actions .

Editing a label is necessary if it inaccurately describes an action that did not occur to ensure that the annotations accurately reflect observable reality. This maintains the integrity of the data and prevents misunderstandings or incorrect assumptions about the actions documented .

An error in object identification, such as mislabeling the color of an object ('orange box' instead of 'yellow box'), must be corrected because it affects the accuracy and reliability of the annotation. Correcting such errors ensures that the annotations precisely reflect the observed reality, facilitating proper understanding and analysis of the captured data .

Consistency in terminology, such as using specific terms like 'lift', 'pick up', and 'grab', improves the quality of annotations by reducing ambiguity and ensuring clarity. Consistent terminology allows for better comparison and interpretation of the data across different instances .

Annotations must capture all actions that involve meaningful object interactions when multiple events happen simultaneously. It is important to not just capture the first or most important action, but to document all significant interactions that occur .

You might also like