Activity Specification

Author: Lijun Yu

Email: [email protected]

Submodule for specification and evaluation of activities in the ActEV context.

Activity Format

Cube Activities

Spatial-temporal cube activities in a video, with optional confidence scores. Each activity is represented as (id, type, score, t0, t1, x0, y0, x1, y1), where type is based on an enum definition.

For class-agnostic proposals,

import torch
from activity_spec import CubeActivities, CubeColumns, ProposalType

cubes = torch.zeros((10, len(CubeColumns)))  # 10 proposals
video_name = 'test.avi'
proposals = CubeActivities(cubes, video_name, ProposalType, CubeColumns)

proposals.to_internal()  # Convert to pandas.DataFrame to view in jupyter

save_dir = '.'
proposals.save(save_dir)  # Save as csv

load_dir = '.'   # Load from csv
proposals = CubeActivities.load(video_name, load_dir, ProposalType)

For activities,

import torch
from activity_spec import ActivityTypes, CubeActivities, CubeColumns

cubes = torch.zeros((10, len(CubeColumns)))  # 10 activities
video_name = 'test.avi'
dataset = 'MEVA'
cube_acts = CubeActivities(
    cubes, video_name, ActivityTypes[dataset], CubeColumns)

cube_acts.to_official()  # Convert to official Json structure

For wrapped labels,

import torch
from activity_spec import ActivityTypes, CubeActivities

video_name = 'test.avi'
dataset = 'MEVA'
labels = torch.zeros((10, len(ActivityTypes[dataset])))
wrapped_labels = CubeActivities(labels, video_name, None, ActivityTypes[dataset])

save_dir = '.'
wrapped_labels.save(save_dir)  # Save as csv

load_dir = '.'   # Load from csv
wrapped_labels = CubeActivities.load(video_name, load_dir, None)

See details at cube.py.

Dataset

Generate a clip dataset based on proposals and labels. Number of negative samples can be controlled via the negative_fraction argument.

from activity_spec import ProposalDataset
# Training: random sample negative proposals, drop ignored proposals
train_set = ProposalDataset(file_index_path, proposal_dir, label_dir,
                            video_dir, negative_fraction=1.)
# Testing: use all proposals
test_set = ProposalDataset(file_index_path, proposal_dir, label_dir,
                           video_dir, eval_mode=True)
clip, label = train_set[0]

See details at dataset.py.

Evaluations

ActEV Scorer in Parallel

Run ActEV_Scorer with actev-datasets. The ActEV_Scorer is called in parallel for each type of activity.

cd ..
python -m activity_spec.evaluate <dataset_dir> <subset_name> <prediction_file> <evaluation_dir>
# For example
python -m activity_spec.evaluate \
    actev-datasets/meva \
    kitware_eo_s2-test_99 \
    experiments/xxx/output.json \
    experiments/xxx/eval

Proposal Evaluation

For each proposal generation model, implement Matcher class that matches proposal and ground truth. Then use it with evaluate_proposal.py.

For example,

from activity_spec.evaluate_proposal import main, parse_args

class Matcher(object):

    def __init__(self, args):
        ...  # Initialize something with args

    def __call__(self, cube_acts_ref, cube_acts_det):
        ...  # Match proposals with ground truth and generate labels
        return cube_acts_labeled, wrapped_label_weights

if __name__ == "__main__":
    args = parse_args()
    matcher = Matcher(args)
    main(args, matcher)

Dependency

See actev_base.

License

See License.