ExSample

This is the unofficial project page for the ExSample project. The repository contains the code for the sampling steps in the paper (Algorithm 1). More technical details can be find in the paper ExSample: Efficient Searches on Video Repositories through Adaptive Sampling.

Requirements

You will need the following installed:

pip install -r requirements.txt

This repository uses swag to access video clips.

git clone https://github.com/stanford-futuredata/swag-python.git
cd swag-python && pip install -e .

Installation

Install this repository with pip local installation.

git clone https://github.com/InkosiZhong/ExSample.git
cd ExSample && pip install -e .

Experiments

Currently, this repository only provides routines for offline query using .csv annotation files. If you want to query online with your object detection NN and tracking algorithms, please inherit class BaseQuery in ExSample/query.py and implement related functions.

Data Preparation

This repository borrowed the video clips and .csv annotation files from stanford-futuredata.

1. The files can be downloaded from the Google Drive.

2. Modify -v and -c in the ExSample/examples/archie.sh.

Run Routine

This routine will query bicycle objects in the Archie dataset with chunks=16 and limit=3000.

sh ExSample/examples/archie.sh

The script will output the objects, frame indices and other metrics.

frame-id:
[0] 731072 [1] 439370 [2] 417376 ...

objects:
[0] Object(label=bicycle bbox=Bbox(2280, 1865, 2538, 2000) conf=1.00 track_id=266371)
[1] Object(label=bicycle bbox=Bbox(2361, 1873, 2596, 2010) conf=0.94 track_id=160599)
[2] Object(label=bicycle bbox=Bbox(2020, 1958, 2287, 2105) conf=0.99 track_id=152919)
...

ExSample Offline Query
------------------------  --
nb_find                    3000
nb_samples                 8673

Custom Query

Check the following arguments to run a custom query on your dataset:

usage: offline_query.py [-h] --video-path VIDEO_PATH --csv-path CSV_PATH
                        --query-objects QUERY_OBJECTS [QUERY_OBJECTS ...]
                        [--nb-chunks NB_CHUNKS] --limit LIMIT
                        [--union-find UNION_FIND]

ExSample Offline Query

optional arguments:
  -h, --help            show this help message and exit
  --video-path VIDEO_PATH, -v VIDEO_PATH
                        path of the video
  --csv-path CSV_PATH, -c CSV_PATH
                        path of the csv file
  --query-objects QUERY_OBJECTS [QUERY_OBJECTS ...], -o QUERY_OBJECTS [QUERY_OBJECTS ...]
                        object labels for query
  --nb-chunks NB_CHUNKS, -n NB_CHUNKS
                        number of the chunks, default is 16
  --limit LIMIT, -l LIMIT
                        number of the samples to find
  --union-find UNION_FIND
                        whether to use union-find for discrimination, default
                        is True
  --random-plus RANDOM_PLUS
                        whether to use random+ within Chunk sampling, default 
                        is True