Tango is a research tool for automatically detecting duplicate video-based bug reports by combining visual and textual information present in the videos.

Note: We provide all results. However, for the rankings, we provide all of them except for the visual information models, which we only provide the best performing one (SimCLR-bovw-5fps-1,000vw). We will provide the rest of the rankings for the visual information models upon acceptance


We provide all of our data, models, and results. You can download the zip here: https://drive.google.com/file/d/1VFxzNXlldGDoCkyNeiLrrHCnAPaqJPpH/view?usp=sharing

In this data package, you will find two folders: artifacts and outputs

artifacts contains the videos we collected in our user study, the model files for the different models we evaluated, and the detailed results that we generated (See Detailed Results section for more information). The videos folder is broken down by user, where each user has a folder contain the apps they were given to create videos for. Each of the apps contain folders that denote the bug they generated a report for. Finally, instead these bug folders there is the actual video-based bug report as an mp4 file. The user_assignment.csv just contains the finalized assignments of users to corresponding bug reports.

In the models folder, you will find the two models we evaluated (SIFT, SimCLR, and OCR+IR). In each folder you will find the corresponding trained codebook files that we generated for SIFT and SimCLR. These codebook files are pickle files that contain the binary representation of the trained codebooks. Additionally, in the SimCLR folder, you will find a checkpoint and pytorch model file that contains all the necessary information for reloading our trained SimCLR model. For the OCR+IR folder, you will find all of the code for the OCR+IR model as well as the intermediate output for this particular model, other models' outputs are stored in the outputs folder.

The outputs folder contains all of the intermediate outputs of our code, except for OCR+IR. In the results folder, you will find all of the raw rankings and metrics for the SIFT and SimCLR model for all combinations of video-based bug reports per app. evaluation_setting contains a json file that contains all of the duplicate detection tasks we used for evaluating our models, i.e. setting 2 (See paper for more details). user_rankings_weighted_all and user_results_weighted_all contain converted version of the raw rankings and metrics for the SIFT and SimCLR model to match setting 2. extracted_text contains the output of running the OCR model, i.e. the frames of the videos and the text from each frame. Lastly, combined contains the results of the combined tango approach.

Reproduce Results

We have created our reproduction package using Docker. Please install Docker if you do not already have it install.

Steps to Reproduce:

  1. Download the repository: git clone https://github.com/two-to-tango/tango.git
  2. Download the data from our gdrive: https://drive.google.com/file/d/1VFxzNXlldGDoCkyNeiLrrHCnAPaqJPpH/view?usp=sharing
  3. Unzip the data file
  4. Navigate to the root of the repo and run the start script, passing in the location of the data folder: ./start <data_path>
  5. Once the docker container has finished spinning up, jump into it: docker exec -it tango bash
  6. Change directory into the main folder (cd main) and install tango locally: pip install -e .
  7. Locate the Jupyter Notebook token by running the following command inside the docker container: jupyter notebook list
  8. Open a browser to localhost:8888 and paste the token into the correct field
  9. Navigate to main/nbs/06_results.ipynb and follow the rest of the instructions provided in that notebook

Detailed Results

You can find a spreadsheet containing the results for all of the different configurations we tested at tango_reproduction_package/artifacts/detailed_results.xlsx.

In this excel file, we have multiple sheets. overall shows the performance of the different model configurations averaged across all apps. overall_comb shows the combined performance of the visual and textual model configurations averaged across all apps. Additionally, per-app and per-app-comb has the performance of the single and combined model configurations per app, respectively. Lastly, we provide the overall performance in sheet overall_user_study and overall_user_study_comb of the single and combined model configurations on the settings (used only APOD app) given to the users for evaluating how much time and effort tango can save developers.

All sheets show the performance in terms of mRR (avg_rr), the standard deviation of recipical rank, median (med_rr), and quartile 1 and 3 (q#_rr). The same is true for mAP (avg_ap). We also show the performance in terms of average rank including standard deviation, and quartiles. Lastly, we providing HIT@1-5, 7, and 10 (h#).

Sheets that contain the weight column have information regarding how much weight is given to the visual and textual information. A value of 0.1 means that the textual information received a weight of 0.1 while the visual information was given a weight of 0.9. For values containing two numbers, e.g. 0.1-0.0, refers to the weighting scheme introduced in the paper for when there may be high overlap in vocabulary between duplicate and non-duplicates (See paper for more details). If an app does not have high overlap, then a weight of 0.1 is used for the textual information, else the textual information is not considered, i.e., weight of 0.0.


Download our repo: git clone https://github.com/two-to-tango/tango.git

Navigate to the root of the project and run: pip install .

How to use

tango <query_vid_path> <corpus_path> <codebook_path> <simclr_checkpoint_path>

Example Output:

OrderedDict([   (('APOD', 'CC1', 'U2'), 1.0),
                (('APOD', 'CC1', 'U1'), 0.7101319783997799),
                (('APOD', 'CC3', 'U2'), 0.19613202363022775),
                (('APOD', 'RB', 'U1'), 0.1261138231868554),
                (('APOD', 'CC2', 'U2'), 0.11940588516361622),
                (('APOD', 'CC3', 'U1'), 0.07206934454930929),
                (('APOD', 'CC2', 'U1'), 0.03426603336985401)])

Training SimCLR

For training the SimCLR model we used the RICO dataset and this repository for training a SimCLR model using Pytorch Lightning