Distributed visual attention on a humanoid robot
read more
Citations
CB: A Humanoid Research Platform for Exploring NeuroScience
CB: A Humanoid Research Platform for Exploring NeuroScience
Visual Attention for Robotic Cognition: A Survey
Real-time acoustic source localization in noisy environments for human-robot multimodal interaction
Unconstrained Real-time Markerless Hand Tracking for Humanoid Interaction
References
A feature-integration theory of attention
A model of saliency-based visual attention for rapid scene analysis
Shifts in selective visual attention: towards the underlying neural circuitry.
Early processing of visual information
Computational Neuroscience of Vision
Related Papers (5)
Shifts in selective visual attention: towards the underlying neural circuitry.
Frequently Asked Questions (17)
Q2. What are the future works mentioned in the paper "Distributed visual attention on a humanoid robot" ?
However, the main point the authors wish to make in this paper is that distributed processing is necessary to achieve realtime operation of a complex vision process such as visual attention. The designed architecture can easily scale to accommodate more complex visual processes and the authors intend to use it to implement further visual processes and integrate them together, thus taking a step to a more brainlike processing of visual information on humanoid robots.
Q3. How many features are generated by the above feature processors?
At full resolution (320× 240), the above feature processors generate 2 early feature maps for color (based on double color opponents theory [1]), 1 for intensity, 16 for orientation, 1for motion and 1 for disparity.
Q4. What is the aim of the preattentive mode of visual processing?
The aim of the preattentive mode of visual processing is to select the focus of attention, which is subsequently processed more exhaustively to solve higher-level tasks such as object recognition [4].
Q5. How many features are combined into conspicuity maps?
The combination of feature maps into conspicuity maps for color Ic, intensity Ib, orientation Io, motion Im, and disparities Id involves normalization to a fixed range and searching for global and local maxima to promote feature maps with strong global maxima.
Q6. How does the feature integration theory of attention work?
Feature integration theory of attention postulates that bottom-up preattentive processing is based on exploring the visual search space for various features and integrating them, e. g. by way of saliency maps, until the location of the mostsalient area in the image emerges, e. g. through the competition across various feature maps.
Q7. How fast is the current version of the system?
Human saccadic eye movements are very fast, thus the current version of their eye control system simply moves the robot eyes towards the desired configuration as fast as possible.
Q8. What is the salient area in the image?
The timeintegrated global saliency map is used as an input to a winnertake-all neural network, which is used to compute the most salient area in the image stream (Section II-D).
Q9. What is the difficult task to solve?
The processor that needs to solve the most difficult synchronization task is the one that integrates the conspicuity mapsinto a single saliency map.
Q10. How can the authors scale the architecture to accommodate more complex visual processes?
The designed architecture can easily scale to accommodate more complex visual processes and the authors intend to use it to implement further visual processes and integrate them together, thus taking a step to a more brainlike processing of visual information on humanoid robots.
Q11. What processors are used to transfer the images?
Five of the PCs are equipped with 2×2.2 GHz Intel Xeon processors, two with 2×2.8 GHz Intel Xeon processors, and one with 2 Opteron 250 processors.
Q12. how does the leaky integrate-and-fire model work?
The authors used the leaky integrate-and-fire model to build a three layer 2-D neural network of first order integrators to integrate the contents of the saliency map and choose a focus of attention over time.
Q13. What is the importance of the synchronization of the image streams?
This is essential to make it possible to scale the system to a more advanced vision processing such as shape analysis and object recognition.
Q14. What is the current setup with all the computers connected to a single gigabit switch?
The current setup with all the computers connected to a single gigabit switch proved to be sufficient to transfer the data at full resolutions and frame rates.
Q15. What is the main point of this paper?
Although some of the previous works mention that parallel implementations would be useful and indeed parallel processing was used in at least one of them [9], this is the first study that focuses on issues arising from such a distributed implementation.
Q16. What is the way to split the network?
their implementation of the data transfer routines allows us to split the network into a number of separate networks should the data load become too large.
Q17. What is the way to process conspicuity maps?
Instead of requesting that the data is fully synchronized, the authors monitor the buffer and simultaneously process the data that is as close together in time as possible.