It Takes Two to TANGO: Combining Visual and Textual Information for Detecting Duplicate Video-Based Bug Reports

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. NOTE: SIFT is missing the 10k raw ranking and metrics, but will be provided in a future version. `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.

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`.

To use our data to reproduce our results please visit our github repository: https://github.com/ncoop57/tango