- MILtrack website of Boris Babenko

- problem: tracking an object given just an initial (detection) bounding box

- MILtrack uses an adaptive appearance model, which models not only the object to track, but also the background
- the key idea is to use the Multiple Instance Learning (MIL) paradigm, used before in areas such as object detection + object recognition
- MIL means that during learning, (positive + negative) examples are presented in sets (bags) of image patches, where training labels are provided for sets, rather than individual instances (image patches)
- if a set is labeled positive, it is assumed to contain at least one positive instance, otherwise the bag is labeled negative
- to incorporate MIL into an online tracker, an online MIL version is needed
- it is the first time, that an online MIL algorithm was presented

- detection / tracker location update / applying the classifier to image patches:
- for each new frame a set of image patches around the current tracker location are cropped out
- for each of these patches x, we compute the probability p(y=1|x) and choose the one with the highest probability

- appearance model update:
- two bags are cropped out: a positive bag (with radius < r to the current tracker location) and a negative bag from an annular region (with radius > r and radius < beta)
- the model is then updated using these two bags

- since the features for each weak classifier must be picked a priori, Grabner proposed to use a pool of M candidate weak stump classifiers for choosing the K weak classifiers

- for the offline / batch version of MIL, MILBoost can be used to learn a classifier for image patches
- MILBoost trains a boosting classifier that maximizes the (log) likelihood of bags
- the probability of a bag being positiv is expressed in terms of the probabilities of its instances being positive, using the so called (Noisy-)OR model

- Oza developed an online variant of the popular AdaBoost (offline/batch) algorithm
- for an incoming example x, each weak classifier is updated sequentially and the weight of example x is adjusted after each update
- since the update formulas for the example weights and classifer weights in AdaBoost depend only on the error of the weak classifiers, we can use a running average of the error for each weak classifier, to estimate the example + classifier weight in an online manner

- all weak classifiers if the pool (containing M classifiers) are updated in parallel
- then K weak classifiers h are chosen from the candidate pool sequentially by maximizing the log likelihood of the bags

Here are 3 own experiments to track objects / subjects using MILtrack:

Car / Toy tracking:

Bike tracking / KITTI dataset:

Football player tracking:

public/multiple_instance_learning_tracking.txt · Last modified: 2013/09/13 09:40 (external edit) · []