Title: UMD Evaluation of Object Tracking in Video
1UMD Evaluation of Object Tracking in Video
2VACE Project
- Multiple teams from different research groups
present algorithms for different video problems. - Evaluation is handled by several teams
- Penn State devises metrics and runs evaluations.
- UMD writes software and authors ground truth.
- ViPER
3ViPER Performance Evaluation
- Java program for support at UMD.
- Work started in mid-1990s.
- Modified to support Penn State metrics in 2000.
4Penn State Frame Evaluations
- Look at the results for each frame, one at a
time. - For each frame, do some set of evaluation
metrics. These include - Object count precision and recall.
- Pixel precision and recall over all objects in
frame. - Individual object pixel precision and recall
measures.
5Penn State Tracking Retrieval
- Assumes matching has already been accomplished.
- First Frame
- For the life of the object, calculate some set of
metrics. - Gets a set of distances for each frame.
- Can display as a line graph with time.
- Can get an average for each distance.
6But
- Frame metrics throw away tracking information.
- Tracking metrics require a known matching.
- This constraint alters the problem.
- Even with the known matching, does not handle
tracking adequately, to include things like
confusion and occlusion. - The results, as described, are just sums over all
frames. - There is no unified metric across time and space.
7UMD Maximal Optimal Matching
- Score for each possible object match.
- Find the optimal matching.
- One-to-one Match Get the list of pairs that
minimize the total distance over all possible
matchings. - Multiple Match For each disjoint subset of truth
objects, get the disjoint subset of output
objects that minimizes the total overall
distance. - Also get precision and recall.
- For S size of matching
- Precision S / size(candidates)
- Recall S / size(targets)
8UMD Maximal Optimal Matching
- Takes into account both space and time.
- Can be generalized to make no assumptions about
space and time. - Optimal 1-1 matching has many nice properties.
- Can handle many-to-many matching.
9Experimental Results
10Example Tracking Text Object
11Example Tracking Text Frame
12Example Tracking Text Tracking
13Example Person Tracking Object
14Example Person Tracking Frame
15Fin
- Dr. David Doermann
- Dr. Rangachar Kasturi
- David Mihalcik
- Ilya Makedon
- many others
16Tracking Graphs
17Object Level Matching
- Most obvious solution many-many matching.
- Allows matching on any data type, at a price.
18Pixel-Frame-Box Metrics
- Look at each frame and ask a specific question
about its contents. - Number of pixels correctly matched.
- Number of boxes that have some overlap.
- Or overlap greater than some threshold.
- How many boxes overlap a given box?
(Fragmentation) - Look at all frames and ask a question
- Number of frames correctly detected.
- Proper number of objects counted.
19Individual Box Tracking Metrics
- Mostly useful for the retrieval problem, this
solution looks at pairs of ground truth boxes and
a result box. - Metrics are
- Position
- Size
- Orientation
20Questions Ignoring Ground Truth
- Assume the evaluation routine is given a set of
objects to ignore (or rules for determining what
type of object to ignore). How does this effect
the output? - For pixel measures, just dont count pixels on
ignored regions. This works for Tracking and
Frame evaluations. - For object matches, do the complete match when
finished, ignore result data that matches ignored
truth.
21Questions Presenting the Results
- Have some basic built in graphs.
- Line graphs for individual metrics
- Bar charts showing several metrics
- For custom graphs, you have to do it yourself.
- ROC Curves
- Scatter Plots