A sparse object category model for efficient learning and exhaustive recognition
read more
Citations
Object Detection with Discriminatively Trained Part-Based Models
Computer Vision: Algorithms and Applications
LabelMe: A Database and Web-Based Tool for Image Annotation
Image retrieval: Ideas, influences, and trends of the new age
One-shot learning of object categories
References
A Combined Corner and Edge Detector
Visual categorization with bags of keypoints
PCA-SIFT: a more distinctive representation for local image descriptors
Pictorial Structures for Object Recognition
Object class recognition by unsupervised scale-invariant learning
Related Papers (5)
Frequently Asked Questions (13)
Q2. What are the future works in "A sparse object category model for efficient learning and complete recognition" ?
There are several aspects of the model that the authors wish to improve and investigate. Although the authors have restricted the model to a star topology, the approach is applicable to a trees and k-fans [ 4 ], and it will be interesting to determine which topologies are best suited to which type of object category.
Q3. What is the effect of increasing the number of detections/feature-type/image?
Increasing the number of detections/feature-type/image increases the error rate slightly in some cases such as camels, since many of the additional detections lie in the background of the image, so increasing the chances of a false positive.
Q4. What is the second method to learn the HSM?
The second method, which the authors adopt, is to learn the HSM directly using EM as in [8, 21], starting from randomly-chosen initial conditions, enabling the learning of many more parts and with more detections/image.
Q5. What are the main advantages of using feature-based methods?
The majority of approaches using feature-based methods rely on region detectors such as Kadir and Brady or multi-scale Harris [11, 13] which favour interest points or circular regions.
Q6. What is the method to learn a heterogeneous star model?
One method is to learn a fully connected constellation model using EM [8] and then reduce the learnt spatial model to a star by completely trying out each of the parts as a landmark, and picking the one which gives the highest likelihood on the training data.
Q7. What is the limitation of learning a heterogeneous star model?
The limitation of this approach is that the fully connected model can only handle a small number of parts and detections in learning.
Q8. What is the cost of finding the optimal match?
The authors then introduce the shape model, which by the use of distance transforms [6], reduces the cost of finding the optimal match from O(N 2P ) to O(NP ).
Q9. How many datasets are used to evaluate the HSM?
Evaluation of the HSM using feature-based detection is done using nine widely varying, unnormalized, datasets summarized in Table 1.
Q10. What was the first evidence that the object category could be learned from weaklysupervised training images?
The constellation model [3, 8, 21] was the first to convincingly demonstrate that models could be learnt from weaklysupervised unsegmented training images (i.e. the only supervision information was that the image contained an instance of the object category, but not the location of the instance in the image).
Q11. What is the problem with using a large number of regions?
To ensure this, one approach is to use a very large number of regions, however the problem remains that each feature will still be perturbed slightly in location and scale from its optimal position so degrading the quality of the match obtainable by the model.
Q12. What is the gradient of the star model?
The star model’s curve, while also roughly linear, has a much flatter gradient: a 12 part star model taking the same time to learn as a 6 part full model.
Q13. What is the likelihood of the hypothesis?
The learnt model is then applied to the regions/curves and the likelihood of the best hypothesis computed using the learnt model.