Skip to content

results

sleap.qc.results

Result classes for Label QC.

Classes:

Name Description
FrameKey

Key to identify a specific frame in a Labels object.
FrameQC

Quality check results for a single frame.
InstanceKey

Key to identify a specific instance in a Labels object.
QCFlag

Single flagged instance with explanation.
QCResults

Container for all QC results.

FrameKey

Bases: NamedTuple

Key to identify a specific frame in a Labels object.

Source code in sleap/qc/results.py 
21
22
23
24
25
class FrameKey(NamedTuple):
    """Key to identify a specific frame in a Labels object."""

    video_idx: int
    frame_idx: int

FrameQC dataclass

Quality check results for a single frame.

Source code in sleap/qc/results.py 
28
29
30
31
32
33
34
35
36
37
@dataclass
class FrameQC:
    """Quality check results for a single frame."""

    is_incomplete: bool = False
    expected_instance_count: int = 0
    actual_instance_count: int = 0
    duplicate_pairs: list[tuple[int, int]] = field(default_factory=list)
    duplicate_reasons: list[str] = field(default_factory=list)
    is_negative_with_instances: bool = False

InstanceKey

Bases: NamedTuple

Key to identify a specific instance in a Labels object.

Source code in sleap/qc/results.py 
13
14
15
16
17
18
class InstanceKey(NamedTuple):
    """Key to identify a specific instance in a Labels object."""

    video_idx: int
    frame_idx: int
    instance_idx: int

QCFlag dataclass

Single flagged instance with explanation.

Attributes:

Name Type Description
frame_idx int

Frame index.
instance_idx int

Instance index within the frame.
video_idx int

Video index.
Source code in sleap/qc/results.py 
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
@dataclass
class QCFlag:
    """Single flagged instance with explanation."""

    instance_key: InstanceKey
    score: float
    confidence: str  # "low", "medium", "high"
    top_issue: str
    feature_contributions: dict[str, float]
    explanation: str

    @property
    def video_idx(self) -> int:
        """Video index."""
        return self.instance_key.video_idx

    @property
    def frame_idx(self) -> int:
        """Frame index."""
        return self.instance_key.frame_idx

    @property
    def instance_idx(self) -> int:
        """Instance index within the frame."""
        return self.instance_key.instance_idx

frame_idx property

Frame index.

instance_idx property

Instance index within the frame.

video_idx property

Video index.

QCResults dataclass

Container for all QC results.

Attributes:

Name Type Description
instance_scores dict[InstanceKey, float]

Mapping from instance key to anomaly score (0-1).
frame_results dict[FrameKey, FrameQC]

Mapping from frame key to frame-level QC results.
feature_contributions dict[InstanceKey, dict[str, float]]

Mapping from instance key to per-feature scores.
feature_names list[str]

List of feature names used.

Methods:

Name Description
get_explanation

Get human-readable explanation for instance.
get_flagged

Get instances flagged above threshold.
get_frame_issues

Get frames with issues (incomplete, duplicates, or bad negatives).
to_dataframe

Export results as DataFrame.
Source code in sleap/qc/results.py 
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
@dataclass
class QCResults:
    """Container for all QC results.

    Attributes:
        instance_scores: Mapping from instance key to anomaly score (0-1).
        frame_results: Mapping from frame key to frame-level QC results.
        feature_contributions: Mapping from instance key to per-feature scores.
        feature_names: List of feature names used.
    """

    instance_scores: dict[InstanceKey, float] = field(default_factory=dict)
    frame_results: dict[FrameKey, FrameQC] = field(default_factory=dict)
    feature_contributions: dict[InstanceKey, dict[str, float]] = field(
        default_factory=dict
    )
    feature_names: list[str] = field(default_factory=list)

    def get_flagged(self, threshold: float = 0.7) -> list[QCFlag]:
        """Get instances flagged above threshold.

        Args:
            threshold: Score threshold (0-1). Instances with scores >= threshold
                are flagged.

        Returns:
            List of QCFlag objects, sorted by score descending.
        """
        flagged = []
        for key, score in self.instance_scores.items():
            if score >= threshold:
                contributions = self.feature_contributions.get(key, {})
                top_issue = self._infer_top_issue(contributions)
                confidence = self._get_confidence(score, contributions)
                explanation = self._generate_explanation(
                    score, top_issue, contributions
                )

                flagged.append(
                    QCFlag(
                        instance_key=key,
                        score=score,
                        confidence=confidence,
                        top_issue=top_issue,
                        feature_contributions=contributions,
                        explanation=explanation,
                    )
                )

        # Sort by score descending
        flagged.sort(key=lambda f: f.score, reverse=True)
        return flagged

    def get_frame_issues(self) -> list[tuple[FrameKey, FrameQC]]:
        """Get frames with issues (incomplete, duplicates, or bad negatives)."""
        issues = []
        for key, frame_qc in self.frame_results.items():
            if (
                frame_qc.is_incomplete
                or frame_qc.duplicate_pairs
                or frame_qc.is_negative_with_instances
            ):
                issues.append((key, frame_qc))
        return issues

    def get_explanation(self, instance_key: InstanceKey) -> str:
        """Get human-readable explanation for instance."""
        score = self.instance_scores.get(instance_key)
        if score is None:
            return "Instance not found in results."

        contributions = self.feature_contributions.get(instance_key, {})
        top_issue = self._infer_top_issue(contributions)
        return self._generate_explanation(score, top_issue, contributions)

    def to_dataframe(self) -> "pd.DataFrame":
        """Export results as DataFrame.

        Returns:
            DataFrame with columns: video_idx, frame_idx, instance_idx, score,
            confidence, top_issue, and one column per feature.
        """
        import pandas as pd

        rows = []
        for key, score in self.instance_scores.items():
            contributions = self.feature_contributions.get(key, {})
            row = {
                "video_idx": key.video_idx,
                "frame_idx": key.frame_idx,
                "instance_idx": key.instance_idx,
                "score": score,
                "confidence": self._get_confidence(score, contributions),
                "top_issue": self._infer_top_issue(contributions),
            }
            row.update(contributions)
            rows.append(row)

        return pd.DataFrame(rows)

    def _infer_top_issue(self, contributions: dict[str, float]) -> str:
        """Infer the most likely issue from feature contributions.

        Normalizes contributions to comparable scales before finding the
        dominant feature, since z-score features (~0-5) and raw distance
        features (~0-100+) have different magnitudes.
        """
        if not contributions:
            return "Unknown"

        # Normalize contributions to comparable scales
        # Z-score features are already ~0-5 range, raw features need scaling
        scale_factors = {
            # Raw distance features - scale to ~0-5 range
            "max_centroid_distance": 30.0,
            "centroid_distance_std": 10.0,
            "nn_distance": 10.0,
            # Curvature is typically 0-3
            "max_curvature": 1.0,
            "curvature_std": 1.0,
            # Rate features (0-1 range) - scale up to be comparable
            "visibility_rate": 0.3,
            "visibility_pattern_score": 0.3,
            "has_isolated_invisible": 0.3,
            # Symmetry: only meaningful if skeleton has symmetry defined
            # Value of 1.0 usually means no symmetry info, so scale down
            "min_symmetry_consistency": 5.0,
        }

        normalized = {}
        for feat, val in contributions.items():
            scale = scale_factors.get(feat, 1.0)
            # Skip features with default/uninformative values
            if feat == "min_symmetry_consistency" and val == 1.0:
                normalized[feat] = 0.0  # Ignore if no symmetry data
            else:
                normalized[feat] = val / scale

        # Find the feature with highest normalized contribution
        top_feature = max(normalized, key=normalized.get)

        # Map feature names to issue descriptions
        issue_map = {
            "max_edge_zscore": "Unusual edge length",
            "mean_edge_zscore": "Unusual proportions",
            "max_angle_zscore": "Unusual joint angle",
            "mean_angle_zscore": "Unusual pose structure",
            "max_pairwise_zscore": "Unusual node spacing",
            "mean_pairwise_zscore": "Unusual scale",
            "bbox_area_zscore": "Unusual scale",
            "max_centroid_distance": "Isolated node",
            "centroid_distance_std": "Inconsistent spacing",
            "min_symmetry_consistency": "Likely L/R swap",
            "visibility_rate": "Unusual visibility",
            "has_isolated_invisible": "Isolated invisible node",
            "visibility_pattern_score": "Unusual visibility pattern",
            "nn_distance": "Unusual pose shape",
            "max_curvature": "Unusual curvature",
            "hull_area_zscore": "Unusual pose extent",
        }

        return issue_map.get(top_feature, f"High {top_feature}")

    def _get_confidence(self, score: float, contributions: dict[str, float]) -> str:
        """Determine confidence level."""
        if score > 0.8:
            return "high"
        elif score > 0.5:
            return "medium"
        return "low"

    def _generate_explanation(
        self, score: float, top_issue: str, contributions: dict[str, float]
    ) -> str:
        """Generate human-readable explanation."""
        lines = [f"Anomaly score: {score:.2f}", f"Primary issue: {top_issue}"]

        if contributions:
            # Get top 3 contributing features
            sorted_features = sorted(
                contributions.items(), key=lambda x: x[1], reverse=True
            )[:3]
            lines.append("Top contributing features:")
            for feature, value in sorted_features:
                lines.append(f"  - {feature}: {value:.3f}")

        return "\n".join(lines)

get_explanation(instance_key)

Get human-readable explanation for instance.

Source code in sleap/qc/results.py 
132
133
134
135
136
137
138
139
140
def get_explanation(self, instance_key: InstanceKey) -> str:
    """Get human-readable explanation for instance."""
    score = self.instance_scores.get(instance_key)
    if score is None:
        return "Instance not found in results."

    contributions = self.feature_contributions.get(instance_key, {})
    top_issue = self._infer_top_issue(contributions)
    return self._generate_explanation(score, top_issue, contributions)

get_flagged(threshold=0.7)

Get instances flagged above threshold.

Parameters:

Name Type Description Default
threshold float

Score threshold (0-1). Instances with scores >= threshold are flagged.

 0.7

Returns:

Type Description
list[QCFlag]

List of QCFlag objects, sorted by score descending.
Source code in sleap/qc/results.py 
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
def get_flagged(self, threshold: float = 0.7) -> list[QCFlag]:
    """Get instances flagged above threshold.

    Args:
        threshold: Score threshold (0-1). Instances with scores >= threshold
            are flagged.

    Returns:
        List of QCFlag objects, sorted by score descending.
    """
    flagged = []
    for key, score in self.instance_scores.items():
        if score >= threshold:
            contributions = self.feature_contributions.get(key, {})
            top_issue = self._infer_top_issue(contributions)
            confidence = self._get_confidence(score, contributions)
            explanation = self._generate_explanation(
                score, top_issue, contributions
            )

            flagged.append(
                QCFlag(
                    instance_key=key,
                    score=score,
                    confidence=confidence,
                    top_issue=top_issue,
                    feature_contributions=contributions,
                    explanation=explanation,
                )
            )

    # Sort by score descending
    flagged.sort(key=lambda f: f.score, reverse=True)
    return flagged

get_frame_issues()

Get frames with issues (incomplete, duplicates, or bad negatives).

Source code in sleap/qc/results.py 
120
121
122
123
124
125
126
127
128
129
130
def get_frame_issues(self) -> list[tuple[FrameKey, FrameQC]]:
    """Get frames with issues (incomplete, duplicates, or bad negatives)."""
    issues = []
    for key, frame_qc in self.frame_results.items():
        if (
            frame_qc.is_incomplete
            or frame_qc.duplicate_pairs
            or frame_qc.is_negative_with_instances
        ):
            issues.append((key, frame_qc))
    return issues

to_dataframe()

Export results as DataFrame.

Returns:

Type Description
'pd.DataFrame'

DataFrame with columns: video_idx, frame_idx, instance_idx, score, confidence, top_issue, and one column per feature.
Source code in sleap/qc/results.py 
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
def to_dataframe(self) -> "pd.DataFrame":
    """Export results as DataFrame.

    Returns:
        DataFrame with columns: video_idx, frame_idx, instance_idx, score,
        confidence, top_issue, and one column per feature.
    """
    import pandas as pd

    rows = []
    for key, score in self.instance_scores.items():
        contributions = self.feature_contributions.get(key, {})
        row = {
            "video_idx": key.video_idx,
            "frame_idx": key.frame_idx,
            "instance_idx": key.instance_idx,
            "score": score,
            "confidence": self._get_confidence(score, contributions),
            "top_issue": self._infer_top_issue(contributions),
        }
        row.update(contributions)
        rows.append(row)

    return pd.DataFrame(rows)