Semi-supervised person re-identification using multi-view clustering
A semi-supervised model incorporates deep feature learning and pseudo label estimation into a unified framework. The deep feature learning can include multiple convolutional neural networks (CNNs). The CNNs can be trained on available training datasets, tuned using a small amount of labeled training samples, and stored as the original models. Features are then extracted for unlabeled training samples by utilizing the original models. Multi-view clustering is used to cluster features to generate pseudo labels. Then the original models are tuned by using an updated training set that includes labeled training samples and unlabeled training samples with pseudo labels. Iterations of multi-view clustering and tuning using an updated training set can continue until the updated training set is stable.
1. A computer system, comprising:
a set of person re-identification neural networks;
a processor; and
system memory coupled to the processor and storing instructions configured to cause the processor to:
initialize the set of person re-identification neural networks on labeled training data;
iteratively tune the set of person re-identification neural networks, including:
extract human facial features from an unlabeled data set;
perform multi-view clustering on the extracted human facial features estimating pseudo labels for the unlabeled data set; and
tune the set of person re-identification neural networks using the unlabeled data along with the estimated pseudo labels improving person re-identification accuracy of the set of person re-identification neural networks, after tuning:
extract other human facial features from the unlabeled data set;
perform multi-view clustering on the other extracted human facial features estimating other pseudo labels for the unlabeled data set; and
further tune the set of person re-identification neural networks using the unlabeled data along with the estimated other pseudo labels further improving person re-identification accuracy of the set of person re-identification neural networks;
wherein instructions configured to perform multi-view clustering comprise instructions configured to calculate an identification loss and a ranking loss per extracted human facial feature.
2. The computer system of claim 1 , wherein the set of neural networks is a set of convolutional neural networks.
3. The neural network architecture of claim 1 , wherein instructions configured to initialize the set of neural networks on labeled training data comprises instructions configured to initialize the set of neural networks on ImageNet data.
4. A method comprising:
initializing a set of person re-identification neural networks on labeled training data;
iteratively tuning the set of person re-identification neural networks, including:
extracting human facial features from an unlabeled data set;
performing multi-view clustering on the extracted human facial features estimating pseudo labels for the unlabeled data set; and
tuning the set of person re-identification neural networks into a tuned set of person re-identification neural networks using the unlabeled data along with the estimated pseudo labels improving person re-identification accuracy of the set of person re-identification neural networks; after tuning:
extracting other human facial features from the unlabeled data set;
performing multi-view clustering on the other extracted human facial features estimating other pseudo labels for the unlabeled data set; and
further tuning the tuned set of person re-identification neural networks using the unlabeled data along with the estimated other pseudo labels further improving person re-identification accuracy of the set of person re-identification neural networks;
wherein performing multi-view clustering comprises calculating an identification loss and a ranking loss per extracted human facial feature.
5. The method of claim 4 , wherein the set of neural networks is a set of convolutional neural networks.
6. The method of claim 4 , wherein initializing the set of neural networks on labeled training data comprises initializing the set of neural networks on ImageNet data.
7. The method of claim 4 , wherein extracting human facial features from an unlabeled data set comprises the set of person re-identification neural networks extracting the human facial features from the unlabeled data set; and
wherein extracting other human facial features from the unlabeled data set comprises the set of tuned person re-identification neural networks extracting the other human facial features from the unlabeled data set.
8. The method of claim 4 , wherein estimating pseudo labels for the unlabeled data set comprises calculating a centroid matrix for the extracted human facial features and based on a weight factor and weight distribution control parameter, the centroid matrix satisfying a coding scheme and being an expected number of clusters.
9. The method of claim 4 , further comprising updating the extracted human facial features with the estimated pseudo labels.
10. The method of claim 4 , wherein estimating other pseudo labels for the unlabeled data set comprises re-estimating the pseudo labels; and,
wherein further turning the tuned set of person re-identifier neural networks comprises further tuning the tuned set of person re-identifier neural networks using the re-estimated pseudo labels.