Homography-Based-MOTDT
- Term project results for AAA534 <Computer Vision> in Korea University
- This work is based on MOTDT which is one of the state-of-the-art algorithm for real-time multiple object tracking
- code: https://github.com/longcw/MOTDT
- paper: https://arxiv.org/abs/1809.04427
- For more information, please refer to the report file in this repository
Overview
-
STEP1: Estimate bounding box of frame
t+1
from the current frame t
through Kalman Filter
-
STEP2: Detect object at time
t+1
using R-FCN
-
STEP3: Filter objects estimated in STEP1 and objects detected in STEP2 through Non-Maximum Suppression
-
STEP4: Calculate homography matrix from frame
t
and t+1
-
STEP5: Create candidates by linearly transforming the existing object at time
t
through homography matrix obtained in STEP4
-
STEP6: Allocate bounding box candidates from STEP3 and STEP5 to each object based on IOU and ReIE features.
Tracking Examples
MOT17 Dataset
MOTDT (original)
- The original model cannot maintain the track ID of object 1 (turned to 101), which is covered by object 105
Homography Based MOTDT (proposed)
- Ours maintains the track ID of object 1 and 89 even though they are obscured by object 161 carrying a green bag.
VisDrone Dataset
MOTDT (original)
- The original model cannot maintain the track ID of object 427 (turned to 509) due to a sudden change in camera angle
Homography Based MOTDT (proposed)
- Ours maintains the track ID of object 515 even though there is a sudden change in camera angle at the end of the clip
Results
MOT17 Dataset
|
Original |
Proposed |
idf1 |
0.503 |
0.522 |
Mostly Tracked |
59 |
70 |
Mostly Lost |
151 |
152 |
False Positive |
919 |
3,057 |
Num_Misses |
28,580 |
26,781 |
Num_Switches |
200 |
198 |
Num_Fragment |
706 |
574 |
MOTA |
0.428 |
0.421 |
MOTP |
0.152 |
0.164 |
VisDrone Dataset
|
Original |
Proposed |
idf1 |
0.547 |
0.579 |
Mostly Tracked |
75 |
97 |
Mostly Lost |
94 |
97 |
False Positive |
725 |
3,064 |
Num_Misses |
22,704 |
19,818 |
Num_Switches |
504 |
386 |
Num_Fragment |
1,604 |
806 |
MOTA |
0.524 |
0.538 |
MOTP |
0.094 |
0.116 |
Implications
- There has been a clear trade-off between the original and proposed method
-
False Positive increased a lot with additional bounding boxes generated by Homography, while Mostly Tracked measure which means the tracking success in the 80% of whole frames improved
- Additionally, number of misses and number of fragments decreased considerably because of supplementary bounding boxes
-
Tracking time increased enormously, which is main downside of proposed method