Spaces:
Sleeping
title: panoptic-quality
tags:
- evaluate
- metric
description: PanopticQuality score
sdk: gradio
sdk_version: 4.44.1
app_file: app.py
pinned: false
emoji: 🖼️
SEA-AI/PanopticQuality
This hugging face metric uses seametrics.segmentation.PanopticQuality under the hood to compute a panoptic quality score. It is a wrapper class for the torchmetrics class torchmetrics.detection.PanopticQuality.
Getting Started
To get started with PanopticQuality, make sure you have the necessary dependencies installed. This metric relies on the evaluate, seametrics and seametrics[segmentation]libraries for metric calculation and integration with FiftyOne datasets.
Basic Usage
>>> import evaluate
>>> from seametrics.payload.processor import PayloadProcessor
>>> MODEL_FIELD = ["maskformer-27k-100ep"]
>>> payload = PayloadProcessor("SAILING_PANOPTIC_DATASET_QA",
>>> gt_field="ground_truth_det",
>>> models=MODEL_FIELD,
>>> sequence_list=["Trip_55_Seq_2", "Trip_197_Seq_1", "Trip_197_Seq_68"],
>>> excluded_classes=[""]).payload
>>> module = evaluate.load("SEA-AI/PanopticQuality", area_rng=[(0, 100),(100, 1e9)])
>>> module.add_payload(payload, model_name=MODEL_FIELD[0])
>>> module.compute()
100%|██████████| 3/3 [00:03<00:00, 1.30s/it]
Added data ...
Start computing ...
Finished!
{'scores': {'MOTORBOAT': array([[0. , 0.25889117],
[0. , 0.79029936],
[0. , 0.3275862 ]]),
'FAR_AWAY_OBJECT': array([[0., 0.],
[0., 0.],
[0., 0.]]),
'SAILING_BOAT_WITH_CLOSED_SAILS': array([[0. , 0.35410052],
[0. , 0.75246359],
[0. , 0.47058824]]),
'SHIP': array([[0. , 0.47743301],
[0. , 0.90181785],
[0. , 0.52941179]]),
'WATERCRAFT': array([[0., 0.],
[0., 0.],
[0., 0.]]),
'SPHERICAL_BUOY': array([[0., 0.],
[0., 0.],
[0., 0.]]),
'FLOTSAM': array([[0., 0.],
[0., 0.],
[0., 0.]]),
'SAILING_BOAT_WITH_OPEN_SAILS': array([[0., 0.],
[0., 0.],
[0., 0.]]),
'CONTAINER': array([[0., 0.],
[0., 0.],
[0., 0.]]),
'PILLAR_BUOY': array([[0., 0.],
[0., 0.],
[0., 0.]]),
'AERIAL_ANIMAL': array([[0., 0.],
[0., 0.],
[0., 0.]]),
'HUMAN_IN_WATER': array([[0., 0.],
[0., 0.],
[0., 0.]]),
'WOODEN_LOG': array([[0., 0.],
[0., 0.],
[0., 0.]]),
'MARITIME_ANIMAL': array([[0., 0.],
[0., 0.],
[0., 0.]]),
'WATER': array([[0. , 0.96737861],
[0. , 0.96737861],
[0. , 1. ]]),
'SKY': array([[0. , 0.93018024],
[0. , 0.93018024],
[0. , 1. ]]),
'LAND': array([[0. , 0.53552331],
[0. , 0.84447907],
[0. , 0.63414633]]),
'CONSTRUCTION': array([[0., 0.],
[0., 0.],
[0., 0.]]),
'OWN_BOAT': array([[0., 0.],
[0., 0.],
[0., 0.]]),
'ALL': array([[0. , 0.18544773],
[0. , 0.27297993],
[0. , 0.20851224]])},
'numbers': {'MOTORBOAT': array([[ 0. , 19. ],
[ 6. , 18. ],
[10. , 60. ],
[ 0. , 15.01568782]]),
'FAR_AWAY_OBJECT': array([[0., 0.],
[6., 6.],
[9., 0.],
[0., 0.]]),
'SAILING_BOAT_WITH_CLOSED_SAILS': array([[0. , 4. ],
[0. , 6. ],
[0. , 3. ],
[0. , 3.00985438]]),
'SHIP': array([[ 0. , 9. ],
[ 0. , 2. ],
[ 1. , 14. ],
[ 0. , 8.11636066]]),
'WATERCRAFT': array([[ 0., 0.],
[ 1., 9.],
[11., 1.],
[ 0., 0.]]),
'SPHERICAL_BUOY': array([[ 0., 0.],
[ 1., 3.],
[36., 0.],
[ 0., 0.]]),
'FLOTSAM': array([[0., 0.],
[0., 0.],
[7., 4.],
[0., 0.]]),
'SAILING_BOAT_WITH_OPEN_SAILS': array([[0., 0.],
[0., 5.],
[0., 0.],
[0., 0.]]),
'CONTAINER': array([[0., 0.],
[0., 0.],
[0., 0.],
[0., 0.]]),
'PILLAR_BUOY': array([[0., 0.],
[0., 0.],
[5., 3.],
[0., 0.]]),
'AERIAL_ANIMAL': array([[0., 0.],
[0., 0.],
[0., 0.],
[0., 0.]]),
'HUMAN_IN_WATER': array([[0., 0.],
[0., 0.],
[0., 0.],
[0., 0.]]),
'WOODEN_LOG': array([[0., 0.],
[0., 0.],
[0., 0.],
[0., 0.]]),
'MARITIME_ANIMAL': array([[0., 0.],
[0., 0.],
[0., 0.],
[0., 0.]]),
'WATER': array([[ 0. , 24. ],
[ 0. , 0. ],
[ 0. , 0. ],
[ 0. , 23.21708667]]),
'SKY': array([[ 0. , 24. ],
[ 0. , 0. ],
[ 0. , 0. ],
[ 0. , 22.32432568]]),
'LAND': array([[ 0. , 13. ],
[ 0. , 7. ],
[ 0. , 8. ],
[ 0. , 10.97822797]]),
'CONSTRUCTION': array([[0., 0.],
[0., 0.],
[0., 0.],
[0., 0.]]),
'OWN_BOAT': array([[ 0., 0.],
[ 0., 0.],
[ 0., 10.],
[ 0., 0.]]),
'ALL': array([[ 0. , 93. ],
[ 14. , 56. ],
[ 79. , 103. ],
[ 0. , 82.66154319]])}}
Metric Settings
The metric takes four optional input parameters: label2id, stuff, per_class, split_sq_rq, area_rng, class_agnostic and method.
label2id: Dict[str, int]: this dictionary is used to map string labels to an integer representation. if not provided a default setting will be used:{'WATER': 0, 'SKY': 1, 'LAND': 2, 'MOTORBOAT': 3, 'FAR_AWAY_OBJECT': 4, 'SAILING_BOAT_WITH_CLOSED_SAILS': 5, 'SHIP': 6, 'WATERCRAFT': 7, 'SPHERICAL_BUOY': 8, 'CONSTRUCTION': 9, 'FLOTSAM': 10, 'SAILING_BOAT_WITH_OPEN_SAILS': 11, 'CONTAINER': 12, 'PILLAR_BUOY': 13, 'AERIAL_ANIMAL': 14, 'HUMAN_IN_WATER': 15, 'OWN_BOAT': 16, 'WOODEN_LOG': 17, 'MARITIME_ANIMAL': 18}stuff: List[str]: this list holds all string labels that belong to stuff. if not provided a default setting will be used:["WATER", "SKY", "LAND", "CONSTRUCTION", "ICE", "OWN_BOAT"]per_class: bool = True: By default, the results are split up per class. Setting this to False will aggregate the results (average the scores, sum up the numbers; see below for explanation of scores and numbers)split_sq_rq: bool = True: By default, the PQ-score is returned in three parts: the PQ score itself, and split into the segmentation quality (SQ) and recognition quality (RQ) part. Setting this to False will return the PQ score only (PQ=RQ*SQ).area_rng: List[Tuple[float]]: The list holds all the area ranges for which results are calculated. Each range is represented by a Tuple, where the first element is the lower limit and the second is the upper limit of the area range. Each value represents total number of pixels of a mask. The parameter defaults to [(0, 1e5 ** 2),(0 ** 2, 6 ** 2),(6 ** 2, 12 ** 2),(12 ** 2, 1e5 ** 2)].class_agnostic: bool = False: If true, all instance labels will be merged to a single instance class (class agnostic), while semantic classes are preserved.method: Litera["iou", "hungarian"] = "hungarian": Controls the method used to match predictions to ground truths. If "iou", then a prediction is matched with a ground truth if IOU > 0.5 (https://arxiv.org/pdf/1801.00868). Can lead to unintuitive results. If "hungarian", then predictions are matched with a ground truth by an hungarian optimizer, which allows also matches with 0 < iou <= 0.5 (https://arxiv.org/abs/2309.04887). Both methods result in a one-to-one mapping.
Output Values
A dictionary containing the following keys:
scores: This is a dictionary, that contains a key for each label, if
per_class == True. Otherwise it only contains the key all. For each key, it contains an array that holds the scores in the the rows in following order: PQ, SQ and RQ. Ifsplit_sq_rq == False, the rows consist of PQ only. The number of columns corresponds to the given area ranges. That means, the results in each column are for a certain size of objects.numbers: This is a dictionary, that contains a key for each label, if
per_class == True. Otherwise it only contains the key all. For each key, it contains an array that consists of four elements in the rows: TP, FP, FN and IOU:- TP: number of true positive predictions
- FP: number of false positive predictions
- FN: number of false negative predictions
- IOU: sum of IOU of all TP predictions with ground truth
With all these values, it is possible to calculate the final scores. As for the scores, the number of columns corresponds to the given area ranges. That means, the results in each column are for a certain size of objects.
Further References
- seametrics Library: Explore the seametrics GitHub repository for more details on the underlying library.
- Torchmetrics: https://lightning.ai/docs/torchmetrics/stable/detection/panoptic_quality.html
- Understanding Metrics: The Panoptic Segmentation task, as well as Panoptic Quality as the evaluation metric, were introduced in this paper.
Contribution
Your contributions are welcome! If you'd like to improve SEA-AI/PanopticQuality or add new features, please feel free to fork the repository, make your changes, and submit a pull request.