BlazeIt

This is the official project page for the BlazeIt project.

Please read the paper for full technical details.

Requirements

This repository contains the code for the optimization step in the paper.

You will need the following installed:

• python 3.x
• CUDA, CUDNN
• torch, torchvision, pandas, opencv (with FFMpeg bindings)

Your machine will need at least:

• 300+GB of memory
• 500+GB of space
• A GPU (this has only been tested with NVIDIA P100 and V100)

Installation

You will need to install the following packages:

conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
conda install -c conda-forge opencv
conda install -c conda-forge pyclipper

(conda can be replaced with pip).

You will also need to swag (GitHub here) and blazeit.

Reproducing experiments

IMPORTANT: your runtimes may vary depending on the GPU you are using.

1. Download the data below. The data is expected to be in /lfs/1/ddkang/blazeit/data.

2. Extract the videos into npy files (this is currently required). For exampe, run:

python gen_small_vid.py --base_name jackson-town-square --date 2017-12-14

in the scripts directory.

3. To reproduce the aggregation experiments, in the aggregation folder, run

mkdir csvs
time python run_counter.py --base_name jackson-town-square \
  --train_date 2017-12-14 --thresh_date 2017-12-16 --test_date 2017-12-17 \
  --objects car --no-load_video --out_csv csvs/jackson-town-square-2017-12-17.csv

to generate predicted counts per frame.

Then run

time python run_ebs_sampling.py \
  --obj_name car --err_tol 0.01 \
  --base_name jackson-town-square --test_date 2017-12-17 --train_date 2017-12-14

to run EBS sampling.

4. To reproduce the limit query experiments, in the scrubbing folder, run:

time python taipei-scrubbing.py --base_name jackson-town-square \
  --train_date 2017-12-14 --thresh_date 2017-12-16 --test_date 2017-12-17 \
  --objects car --no-load_video --counts 5 --labeler mock-detectron --limit 10

Datasets

We currently have released the night-street (i.e., jackson-town-square) data. The data is available here.

Please email the first author directly for other datasets.