Skip to content

Latest commit

 

History

History
112 lines (97 loc) · 4.12 KB

README.md

File metadata and controls

112 lines (97 loc) · 4.12 KB

TVR Evalation

Task Definition

Given a natural language query and a large pool of videos (with subtitles), the TVR (VCMR) task requires a system to retrieve a relevant moment from the videos. The table below shows a comparison of the TVR task and the subtasks:

Task Description
VCMR or VSCMR, Video (-Subtitle) Corpus Moment Retrieval. Localize a moment from a large video corpus.
SVMR or SVSMR, Single Video (-Subtitle) Moment Retrieval. Localize a moment from a given video.
VR or VSR, Video (-Subtitle) Retrieval. Retrieve a video from a large video corpus.

VCMR and VR only requires a query and a video corpus, SVMR additionally requires knowing the ground-truth video. Thus it is not possible to perform SVMR on our test-public set, where the ground-truth video is hidden.

How to construct a prediction file?

An example of such file is sample_val_predictions.json, it is formatted as:

{
    "video2idx": {
        "castle_s01e02_seg02_clip_09": 19614,
        ...
    },
    "VCMR": [{
            "desc_id": 90200,
            "desc": "Phoebe puts one of her ponytails in her mouth.",
            "predictions": [
                [19614, 9.0, 12.0, 1.7275],
                [20384, 12.0, 18.0, 1.7315],
                [20384, 15.0, 21.0, 1.7351],
                ...
            ]
        },
        ...
    ],
    "SVMR": [{
            "desc_id": 90200,
            "desc": "Phoebe puts one of her ponytails in her mouth.",
            "predictions": [
                [20092, 36.0, 42.0, -1.9082],
                [20092, 18.0, 24.0, -1.9145],
                [20092, 51.0, 54.0, -1.922],
                ...
            ]
        },
        ...
    ],
    "VR": [{
            "desc_id": 90200,
            "desc": "Phoebe puts one of her ponytails in her mouth.",
            "predictions": [
                [19614, 0, 0, 1.7275],
                [20384, 0, 0, 1.7315],
                [20384, 0, 0, 1.7351],
                ...
            ]
        },
        ...
    ]
}
entry description
video2idx dict, {vid_name: vid_idx}. A mapping of video names to unique video IDs for current set. From tvr_video2dur_idx.json.
VCMR list(dicts), stores predictions for the task VCMR.
SVMR list(dicts), stores predictions for the task SVMR. Not required for test-public submission.
VR list(dicts), stores predictions for the task VR.

The evaluation script will evaluate the predictions for tasks [VCMR, SVMR, VR] independently. Each dict in VCMR/SVMR/VR list is:

{
    "desc": str,
    "desc_id": int,
    "predictions": [[vid_id (int), st (float), ed (float), score (float)], ...]
}

predictions is a list containing 100 sublist, each sublist has exactly 4 items: [vid_id (int), st (float), ed (float), score (float)], which are vid_id (video id), st and ed (moment start and end time, in seconds.), score (score of the prediction). The score item will not be used in the evaluation script, it is left here for record.

Run Evaluation

At project root, run

bash standalone_eval/eval_sample.sh

This command will use eval.py to evaluate the provided sample_val_predictions.json file, the output will be written into sample_val_predictions_metrics.json. Its content should be similar if not the same as sample_val_predictions_metrics_raw.json file.

Codalab Submission

To test your model's performance on test-public set, please submit both val and test-public predictions to our Codalab evaluation server. The submission file should be a single .zip file (no enclosing folder) that contains the two prediction files tvr_test_public_submission.json and tvr_val_submission.json, each of the *submission.json file should be formatted as instructed above. Note that tvr_val_submission.json will have all the 4 entries, while tvr_test_public_submission.json will have only 3 entries, without SVMR.