README.md

Source Explained

For the sake of fast development and submissions, the current module for object detection is not yet updated. Please refer to prediction.py to get the latest developement.

Content

Composed by 1 main process: process_video.py, that takes care of loading the video, setting up the detection_module, the submission_helper, and tracking_module.

• prediction.py: the submission inference src.

• Detection module: is a wrapper of tensorflow implementation, that generate formated prediction.

• Tracking module: generate tracking based on previous-current frame transformation and detection output.

• Submission helper: generate a submission .json file.

• Training label generation.py

• Other experimental scripts...

Tracker

object_tracker.py

Track objects and assign IDs.

stabilizer.py (under development)

Stabilize the camera motion for two adjacent frames and transform the first frame toward the second frame.

Run

from object_tracker import Tracker

...
image_size = (1936, 1216, frame)

tracker = Tracker(image_size)
...

# input prediction data for each frame in order
# input: {'box2d': [x1, y1, x2, y1]}
# output: {'id': id, box2d': [x1, y1, x2, y1], 'mv': [vx, vy], 'scale': [sx, sy], 'occlusion': number_of_occlusions}
prediction = tracker.assign_ids(prediction)
...

Test

python3 object_tracker.py --input /path/to/annotation/directory --output /path/to/output.json

Frame #1
    #Boxes: Car=8, Pedestrian=12
    Execution time: total=0.00034165, max=0.00034165
    Total cost:  2e+16
Frame #2
    #Boxes: Car=7, Pedestrian=14
    Execution time: total=0.00176668, max=0.00176668
    Total cost:  2.0000000000000004e+16
...

Frame #600
    #Boxes: Car=5, Pedestrian=14
    Execution time: total=0.00174618, max=0.78549790
    Total cost:  2.43e+18
Overall (../train_annotations/train_02.json)
    Car: total=2984, sw=221, tp=2569, err=0.07406166
    Pedestrian: total=3072, sw=695, tp=1856, err=0.22623698
    All: err=0.15125495