Action recognition using exemplar-based embedding
read more
Citations
A survey on vision-based human action recognition
A survey of vision-based methods for action representation, segmentation and recognition
Human Action Recognition Using Factorized Spatio-Temporal Convolutional Networks
View-Independent Action Recognition from Temporal Self-Similarities
Human Action Recognition using Factorized Spatio-Temporal Convolutional Networks
References
Statistical learning theory
A Decision-Theoretic Generalization of On-Line Learning and an Application to Boosting
An introduction to variable and feature selection
Wrappers for feature subset selection
Visual perception of biological motion and a model for its analysis
Related Papers (5)
Frequently Asked Questions (15)
Q2. How many actors are used to train the classifier?
8 out of the 9 actors in the database are used to train the classifier and select the exemplars, the 9th is used for the evaluation.
Q3. What is the main feature of their approach?
Another important feature of their approach is that it can be used with advanced image matching techniques, such as the Chamfer distance [10], for visual measurements.
Q4. How much recognition rate can the authors achieve with a uniformly sub-sampled exemplar?
For a uniformly sub-sampled exemplar set of size 300, their method presents a recognition rate of 93.6% in crossvalidation on all 10 actions and 9 actors.
Q5. How long does it take to select a set of features?
With a non-optimized implementation in MATLAB, selection of approximately 50 features out of a few hundreds will take around 5 minutes.
Q6. What is the reason for the small size of the training set?
Also note that due to the small size of the training set, the validation rate can easily reach 100% if too many exemplars are considered.
Q7. How much recognition rate do Wang and Suter report?
Wang and Suter [23] report a recognition rate of 97.78% with an approach that uses kernel-PCA for dimensional reduction and factorial conditional random fields to model motion dynamics.
Q8. How can the authors obtain the distance between silhouettes?
Consequently they can be matched with a standard distance function and the authors choose the squared Euclidean distance d(x, y) = |x − y|2, which is computed between the vector representations of the binary silhouette images.
Q9. What is the way to learn and evaluate a candidate exemplar?
in the first iteration classifier for each single candidate exemplar are learned, the exemplar with the best evaluation performance is added to the final exemplar set, and the learning and evaluation step is repeated using pairs of exemplars (containing the already selected), triples, quadruples, etc.
Q10. What is the way to represent a class c 1...C?
Each class c ∈ 1...C is represented through a single Gaussian distribution p(D∗|c) = N (D∗|µc, Σc), which the authors found adequate in experiments to model all important dependencies between exemplars.
Q11. How many exemplars are used in the experiment?
Since the forward selection includes one random step, in the case where several exemplars present the same validation rate, the authors repeat the experiment 10 times with all actors, and average over the results.
Q12. What is the cost of converting distances into probabilities?
Probabilistic exemplar-based approaches [21] do model such uncertainties by converting distances into probabilities, but as mentioned earlier, at the price of complex computations for normalization constants.
Q13. How do the authors compute the Chamfer distance?
An efficient way to compute the Chamfer distance is by correlating the distance transformed observation with the exemplar silhouette.
Q14. What is the case when estimating covariance?
Note that when estimating covariance Σ, and depending on the dimension n, it is often the case that insufficient training data is available for Σ, and consequently the estimation may be non-invertible.
Q15. Why do the authors randomly remove exemplars during the validation step?
In this case, the authors randomly remove exemplars during the validation step, to reduce the validation rate and to allow new exemplars to be added.