THIS REPO IS DEPRECATED. PLEASE USE THE UPDATED VERSION

Caesar-For Edge Computing

How to Use

• Clone this repo to your machine git clone --recurse-submodules https://github.com/USC-NSL/Caesar-Edge.git
• Download the checkpoint models (see below)
• Double check the model path in these files to make sure they are same as yours: object_detector_yolo(#11-13), object_detector_ssd(#11), tracker_deepsort(#10), action_detection_acam(#12)
• Rune pipeline.py

Checkpoint Preparation

All the NN model files should be put into the "checkpoints/" folder. You can download the models in these links:

• SSD
• YOLOv2_cfg, YOLOv2_weights
• DeepSORT
• ACAM
• ReID2

Requirements

See requirements.txt for python packages.

• For SSD, make sure your tensorflow-estimator is compatible with your local tf version. For example, tensorflow-estimator==1.10.12 works for tensorflow-gpu=1.12.0. Otherwise the tf will complain "tf.estimator package not installed"

Components

One module's output will go to the next one

• Video Reader
• Object Detection (SSD, YOLO)
• Tracking (DeepSORT)
• Action Detection (ACAM)

Performance

• SSD512: 25 FPS
• YOLOv2: 22 FPS
• DeepSORT: 49 FPS
• ACAM: 60 FPS/Tube