The Revisor is a computer vision tool that makes elections results transparent and reliable. It automatically counts voter turnout of polling stations allowing civil society to verify official results.

Read our Medium post to learn more about the Revisor's architecture.

Repository contains following modes of the Revisor's core:

Mode name	Purpose	How to enable (edit cfg.py)
Local	Demo and tests	Already enabled
Dataset maker mode	Gather unlabelled actions dataset	Set dataset_maker_mode to True
ReID dataset maker mode	Gather unlabelled ReID dataset	Set reid_dataset_maker_mode to True

Mode for processing huge volumes of data is not presented here («api» mode).

Use case

---

The Revisor was used to process 2018, 2020 and 2021 elections in Russia.

For instance, our team obtained from Meduza 17395 camera videos from 9252 polling stations totalling more than 1 190 000 hours of the 2021 elections.

On acquired data «API» pipeline of the Revisor was launched. Verified counted data of the system showed ~2 million voting actions, meanwhile the official voter turnout was ~3.2 million. It means that we have discovered ~1 million stuffed ballots.

Based on the insights received from these researches Meduza's material was written about falsifications during Russian elections.

How to launch?

---

Preparation

1. Make sure your hardware satisfies requirements and you have all required source data.

2. Install dependencies by following installation guide.

3. Install process manager to run the Revisor without stopping: supervisor, pm2, etc. (optional)

4. Make sure target_dirs.csv rows match with camera folder names located in cfg.source_videos_dir directory (in case if you have custom input videos)

5. Edit cfg.py (in case if you are going to launch dataset maker mode or to enable additional visualization features)

   • Set dataset_maker_mode or reid_dataset_maker_mode to True
   • Set save_labelled_videos to True if you would like to save videos with visualized tracking data
   • Change dataset_dir to your output dataset directory

Launching

1. Go the Revisor's directory in terminal.

2. To process single row of target_dirs.csv (all videos located inside that folder), run:

   python3 revisor.py

   If you launch next iteration, recently processed target directories will be skipped (stored in cfg.processed_videos_csv).

   If the Revisor will be stopped during processing, next iteration will be started from the last unfinished video (stored in cfg.processed_videos_dir)

3. To process all rows of target_dirs.csv, run revisor.py inside process manager. For example, pm2 launching command:

   pm2 start revisor.py --interpreter python3

Outputs

If you want to see the detections and tracking data, enable drawing of cv2 window: set cfg.show_recognitions to True.

All processing results you can find in data/revisor_results directory:

• local datetime of each vote action (stats/votes.csv)
• processed videos statistics (stats/stats.csv)
• samples of the recognized actions (json and videos directories) - specified in The Revisor results section in cfg.py

Improvements

---

We are open to any help. If you want to contribute to the project, we see following things to improve:

• Optimize YOLACT and QueryInst with TensorRT (fp16 quantization) to speed up ballot boxes recognition module
• Move all models to the Triton service
• Call predict from Triton service via gRPC
• Utilize Docker for dependencies installation and launching the Revisor
• Experiment with action recognition models in order to increase accuracy

Our plans

---

• Reveal the Revisor action dataset (ReSet Actions)
• Reveal backbone evolution library (Evolly)
• Reveal re-identification model optimized by Evolly and training pipelines

Contacts

---

Contact us if you are interested in cooperating with us, ready to invest in us or would like to process your elections with the Revisor:

revisorteam@pm.me

References

---

Thanks to following project's authors for their amazing work:

• EvoPose2D
• TensorRT
• TensorRT YOLOv4
• YOLACT
• QueryInst
• SORT tracker
• Centroid tracker