Skip to content


Repository files navigation

Visual Semantic Role Labeling for Video Understanding (CVPR21)

LICENSE Python PyTorch Arxiv

Visual Semantic Role Labeling for Video Understanding
Arka Sadhu, Tanmay Gupta, Mark Yatskar, Ram Nevatia, Aniruddha Kembhavi
CVPR 2021

VidSitu is a large-scale dataset containing diverse 10-second videos from movies depicting complex situations (a collection of related events). Events in the video are richly annotated at 2-second intervals with verbs, semantic-roles, entity co-references, and event relations.

This repository includes:

  1. Instructions to install, download and process VidSitu Dataset.
  2. Code to run all experiments provided in the paper along with log files.
  3. Instructions to submit results to the Leaderboard.


Please see for detailed instructions on downloading and setting up the dataset.


Please see for detailed instructions


  • Basic usage is CUDA_VISIBLE_DEVICES=$GPUS python "experiment_name" --arg1=val1 --arg2=val2 and the arg1, arg2 can be found in configs/vsitu_cfg.yml.

  • Set $GPUS=0 for single gpu training. For multi-gpu training via Pytorch Distributed Data Parallel use $GPUS=0,1,2,3

  • YML has a hierarchical structure which is supported using . For instance, if you want to change the beam_size under gen which in the YML file looks like

        beam_size: 1

    you can pass --gen.beam_size=5

  • Sometimes it might be easier to directly change the default setting in configs/vsitu_cfg.yml itself.

  • To keep the code modular, some configurations are set in code/ as well.

  • All model choices are available under code/

See for detailed usage and reproducing numbers in the paper.


Logs are stored inside tmp/ directory. When you run the code with $exp_name the following are stored:

  • txt_logs/$exp_name.txt: the config used and the training, validation losses after ever epoch.
  • models/$exp_name.pth: the model, optimizer, scheduler, accuracy, number of epochs and iterations completed are stored. Only the best model upto the current epoch is stored.
  • ext_logs/$exp_name.txt: this uses the logging module of python to store the logger.debug outputs printed. Mainly used for debugging.
  • predictions: the validation outputs of current best model.

Logs are also stored using MLFlow. These can be uploaded to other experiment trackers such as, wandb for better visualization of results.

Evaluation (Locally)

  1. Evaluation scripts are available for the three tasks under code/ The same file is used for leaderboard purposes. If you are using this codebase, the predictions are stored under tmp/predictions/{expt_id}/valid_0.pkl. You can evaluate using the following command:

    python code/ --pred_file='./tmp/predictions/{expt_id}/valid_0.pkl' --split_type='valid' --task_type=$TASK

    Here $TASK can be vb, vb_arg, evrel corresponding to Verb Prediction, Semantic Role Prediction and Event Relation Prediction

  2. The output format for the files are as follows:

    1. Verb Prediction:

          # Both lists of length 5. Outer list denotes Events 1-5, inner list denotes Top-5 VerbID predictions
          pred_vbs_ev: List[List[str]]
          # Both lists of length 5. Outer list denotes Events 1-5, inner list denotes the scores for the Top-5 VerbID predictions
          pred_scores_ev: List[List[float]]
          #the index of the video segment used. Corresponds to the number in {valid|test}_split_file.json
          ann_idx: int
    2. Semantic Role Labeling Prediction

          # same as above
          ann_idx: int
          # The main output used for evaluation. Outer Dict is for Events 1-5.
          vb_output: Dict[Dict]
          # The inner dict has the following keys:
              # VerbID of the event
              vb_id: str
              ArgX: str
              ArgY: str

      Note that ArgX, ArgY depend on the specific VerbID

    3. Event Relation Prediction

          # same as above
          ann_idx: int
          # Ouuter list of length 4 and denotes Event Relation {1-3, 2-3, 3-4, 4-5}. Inner list denotes three Event Relations for given Verb+Semantic Role Inputs
          pred_evrels_ev: List[List[str]]
          # Scores for the above
          pred_scores_ev: List[List[float]]

    See examples under docs

Leaderboard (Evaluation on Test Sets)

We maintain three separate leaderboards for each of the three tasks. The leaderboard will accept submissions from April 7th, 2021. The output format remains the same as local evaluation.

Here are the leaderboard links:


          author = {Sadhu, Arka and Gupta, Tanmay and Yatskar, Mark and Nevatia, Ram and Kembhavi, Aniruddha},
          title = {Visual Semantic Role Labeling for Video Understanding},
          booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
          month = {June},
          year = {2021}