Condensed datasets offer a compact representation of larger datasets, but training models directly on them or using them to enhance model performance through knowledge distillation (KD) can result in suboptimal outcomes due to limited information. To address this, we propose a method that expands condensed datasets using model inversion, a technique for generating synthetic data based on the impressions of a pre-trained model on its training data. This approach is particularly well-suited for KD scenarios, as the teacher model is already pre-trained and retains knowledge of the original training data. By creating synthetic data that complements the condensed samples, we enrich the training set and better approximate the underlying data distribution, leading to improvements in student model accuracy during knowledge distillation. Our method demonstrates significant gains in KD accuracy compared to using condensed datasets alone and outperforms standard model inversion-based KD methods by up to 11.4% across various datasets and model architectures. Importantly, it remains effective even when using as few as one condensed sample per class, and can also enhance performance in few-shot scenarios where only limited real data samples are available.

Data-Free Knowledge Distillation (KD) allows knowledge transfer from a trained neural network (teacher) to a more compact one (student) in the absence of original training data. Existing works use a validation set to monitor the accuracy of the student over real data and report the highest performance throughout the entire process. However, validation data may not be available at distillation time either, making it infeasible to record the student snapshot that achieved the peak accuracy. Therefore, a practical data-free KD method should be robust and ideally provide monotonically increasing student accuracy during distillation. This is challenging because the student experiences knowledge degradation due to the distribution shift of the synthetic data. A straightforward approach to overcome this issue is to store and rehearse the generated samples periodically, which increases the memory footprint and creates privacy concerns. We propose to model the distribution of the previously observed synthetic samples with a generative network. In particular, we design a Variational Autoencoder (VAE) with a training objective that is customized to learn the synthetic data representations optimally. The student is rehearsed by the generative pseudo replay technique, with samples produced by the VAE. Hence knowledge degradation can be prevented without storing any samples. Experiments on image classification benchmarks show that our method optimizes the expected value of the distilled model accuracy while eliminating the large memory overhead incurred by the sample-storing methods.

With the increasing popularity of deep learning on edge devices, compressing large neural networks to meet the hardware requirements of resource-constrained devices became a significant research direction. Numerous compression methodologies are currently being used to reduce the memory sizes and energy consumption of neural networks. Knowledge distillation (KD) is among such methodologies and it functions by using data samples to transfer the knowledge captured by a large model (teacher) to a smaller one (student). However, due to various reasons, the original training data might not be accessible at the compression stage. Therefore, data-free model compression is an ongoing research problem that has been addressed by various works. In this paper, we point out that catastrophic forgetting is a problem that can potentially be observed in existing data-free distillation methods. Moreover, the sample generation strategies in some of these methods could result in a mismatch between the synthetic and real data distributions. To prevent such problems, we propose a data-free KD framework that maintains a dynamic collection of generated samples over time. Additionally, we add the constraint of matching the real data distribution in sample generation strategies that target maximum information gain. Our experiments demonstrate that we can improve the accuracy of the student models obtained via KD when compared with state-of-the-art approaches on the SVHN, Fashion MNIST and CIFAR100 datasets.

Model-free object tracking is still challenging because of the limited prior knowledge and the unexpected variation of the target object. In this paper, we propose a feature learning algorithm for model-free multiple object tracking. First, we pre-learn generic features invariant to diverse motion transformations from auxiliary video data by using a deep network of anto-encoder. Then, we adapt the pre-learned features according to multiple target objects respectively in a multi-task learning manner. We treat the feature adaptation for each target object as one single task. We simultaneously learn the common feature shared by all target objects and the individual feature of each object. Experimental results demonstrate that our feature learning algorithm can significantly improve multiple object tracking performance.

In many intelligent surveillance systems there is a requirement to search for people of interest through archived semantic labels. Other than searching through typical appearance attributes such as clothing color and body height, information such as whether a person carries a bag or not is valuable to provide more relevant targeted search. We propose two novel and fast algorithms for sling bag and backpack detection based on the geometrical properties of bags. The advantage of the proposed algorithms is that it does not require shape information from human silhouettes therefore it can work under crowded condition. In addition, the absence of background subtraction makes the algorithms suitable for mobile platforms such as robots. The system was tested with a low resolution surveillance video dataset. Experimental results demonstrate that our method is promising.

Shape and motion are two most distinct cues observed from human actions. Traditionally, K-Nearest Neighbor (K-NN) classifier is used to compute crisp votes from multiple cues separately. The votes are then combined using linear weighting scheme. Usually, the weights are determined in a brute-force or trial-and-error manner. In this study, we propose a new classification framework based on sum-rule fusion of fuzzy K NN classifiers. Fuzzy K-NN classifier is capable of producing soft votes, also known as fuzzy membership values. Based on Bayes theorem, we show that the fuzzy membership values produced by the classifiers can be combined using sum-rule. In our experiment, the proposed framework consistently outperforms the conventional counterpart (K-NN with majority voting) for both Weizmann and KTH datasets. The improvement may attribute to the ability of the proposed framework to handle data ambiguity due to similar poses present in different action classes. We also show that the performance of our method compares favorably with the state-of-the-arts.

Most methods for multiple camera tracking rely on accurate calibration to associate data from multiple cameras. However, it often is not easy to have an accurate calibration in some real applications due to practical reasons. The inaccurate calibration can lead to wrong data association of objects between cameras. In this paper, we propose a method to handle the data association of objects in multiple cameras under inaccurate ground plane homography by using the RFS Bayes filter. Our method is based on modeling measurements from cameras to a random finite set. This random finite set includes the primary measurement from the object, extraneous measurements of the object, and clutter. Experimental results show the efficiency and robustness of our method through challenging cases such as occlusions, merged and split persons.

In practical surveillance applications, methods for multiple camera tracking have to deal with many challenges such as noises and errors in estimating system parameters. In this pa per, we propose a distributed multiple camera tracking system that can track multiple persons in real time under challenging conditions such as inaccuracy from calibration, noises from background and lighting changes. The system combines a network of cameras located around the environment, a central process unit and mobile devices. Experimental results show that our system has already performed well in military training.

Merging and splitting of objects cause challenges for visual tracking. This is due to observation ambiguity, object lost, and tracking errors when objects are close together. In this paper, we propose a method to combine the joint probabilistic data association (JPDA) and the particle filter to maintain tracks of objects. The results of JPDA are employed to improve the observation model in the particle filter. Based on the ability of handling missing detections and clutter of JPDA, tracks of objects can be maintained after merging or splitting. Conversely, the particle filter also improves the performance of JPDA by fusing other observations such as color and background subtraction information. Hence, our method can take advantages from both JPDA and particle filter to track objects through merging and splitting.

Tracking objects using multiple sensors is more efficient than those using one sensor. In this paper, we proposed a method to fuse data from multiple sensors in Gaussian mixture probability hypothesis density filter. This method can avoid the data association problem in multi-sensor multi-object tracking. Moreover, it is more reliable and less computational than particle probability hypothesis density filter for multi-sensor multi-object tracking. We demonstrated the efficient of the approach by applications such as bearing and range tracking, and multiple speaker tracking.

Most methods for multiple object tracking in video represent the state of multi-object in a high dimensional joint state space. This leads to high computational complexity. This paper presents a method using the probability hypothesis density (PHD) filter to estimate the state of multiple objects in video. The method operates on the single object state space instead of the joint state space. A PHD recursion for visual observations with color measurements is proposed. Our method can track varying number of objects.

In this paper, we present an efficient method for tracking multiple speakers in a reverberant environment. The proposed method is based on the cardinalized probability hypothesis density (CPHD) filter. Because the CPHD filter can handle a large amount of clutter measurements, our method has a high reliability when tracking multiple speakers. Simulation experiments are presented to demonstrate the performance of the proposed method.