Adaptive Discriminative Generative Model and Its Applications


This paper presents an adaptive discriminative generative model that generalizes the conventional Fisher Linear Discriminant algorithm and renders a proper probabilistic interpretation. Within the context of object tracking, we aim to find a a discriminative generative model that best separates the target from the background. We present a computationally efficient algorithm to constantly update this discriminative model as time progresses. This web page provides links to the detailed derivation of our algorithm, and the experitmental results of our discriminative generative tracker.

A Discriminative Generative Observation Model

An Adaptive Generative Model

In our algorithm, we model the appearance of the tracked object using probabilistic principle component analysis, and provides a method to online update this generative model. A detailed derivation of online updating the mean and covariance matrix of a PCA can be found in this report .

A Discriminative Generative Model

We then apply discriminative analysis to further disctiguish the tracked object and its background. A detailed derivation of our algorithm to build a discriminative generative model can be found in this report .

Experimental Results

We apply our algorithm to track differnet kinds of objects in indoor and outdoor environment. In the videos of our tracking results, there are two rows of image below the tracking video that show the newly selected examples to update our discriminative generative model. The first row shows the newly selected positive examples and the second row shows the newly selected negative examples in the current frame.

Indoor Sequences

In the indoor sequences, the apperance of the tracked object is constantly changing due to variation of the viewing angle, the illumination condition, and sometimes self-deformation.

1. Tracking a Human Face
In this sequence, our tracker tracks a human face that undergoes large pose variation (moving back and forth between frontal view and profile). The appearance of the track subject continues to change in this video sequence, but our tracker is capable to adapt the change and correctly locate the face thoughtout the sequence.

2. Tracking Sylvester Jr.
In this sequence, the tracked object is a stuff animal, Sylvester Jr. The tracked object also has large pose variation throughout the sequence. Furthermore, the lighting on the tracked object is constantly changing. Under such condition, out tracker still performs well.

3. The Dudek Sequence
This another video of tracking human face. In this sequence, the appearance of the tracked face also undergoes variation due to pose and illumination changes. In addition, there are shift changes in the size and location of the tracked face. Our tracker demonstrates a solid perfromance in this video.

Outdoor Sequences

In ourdoor sequences, the tracker objects undergo shape illumination changes in the video. Also, at times there are irregualar patterns of shadows on appeared on the objects due to part of the sun light blocked by other objects such as tree.

1. Tracking a Car
In this video, we apply our tracker to track a moving car. While there is no apprant pose variation of the tracked object, its appearance does have sudden alteration due to overcast shadow that occurs when the car pass by trees or pass beneath a bridge. The experimental result shows how our tracker reacts to these changes.

2. Tracking a Pedestrian
In this video, we track the upper torso of a moving pedestrian. The pedestrain moves from the shadow into sunlight, so there is a keen illumination change. In addition, in order to track the whole upper torso, our bounding window inevitably contains a lot of background pixels, which might deteriorates the perfromance of a appearance-based tracker. However, our tracker still works fine.