File size: 7,458 Bytes
f60a6c1 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 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 65 66 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 |
from logging import getLogger
from scorevision.vlm_pipeline.domain_specific_schemas.challenge_types import (
ChallengeType,
parse_challenge_type,
)
from scorevision.utils.data_models import (
SVChallenge,
SVRunOutput,
SVEvaluation,
TotalScore,
)
from scorevision.chute_template.schemas import TVPredictInput
from scorevision.vlm_pipeline.non_vlm_scoring.keypoints import evaluate_keypoints
from scorevision.vlm_pipeline.non_vlm_scoring.objects import (
compare_object_counts,
compare_team_labels,
compare_object_labels,
compare_object_placement,
)
from scorevision.utils.settings import get_settings
from scorevision.utils.video_processing import FrameStore
from scorevision.vlm_pipeline.utils.data_models import (
PseudoGroundTruth,
MinerScore,
AggregatedScore,
)
from scorevision.vlm_pipeline.utils.response_models import (
FrameAnnotation,
BoundingBox,
ShirtColor,
TEAM1_SHIRT_COLOUR,
TEAM2_SHIRT_COLOUR,
)
from scorevision.vlm_pipeline.domain_specific_schemas.football import (
Person as ObjectOfInterest,
OBJECT_ID_LOOKUP,
)
from scorevision.vlm_pipeline.domain_specific_schemas.football import Action
from scorevision.vlm_pipeline.non_vlm_scoring.smoothness import bbox_smoothness_per_type
logger = getLogger(__name__)
def parse_miner_prediction(miner_run: SVRunOutput) -> dict[int, dict]:
predicted_frames = (
(miner_run.predictions or {}).get("frames") if miner_run.predictions else None
) or []
logger.info(f"Miner predicted {len(predicted_frames)} frames")
miner_annotations = {}
for predicted_frame in predicted_frames:
bboxes = []
frame_number = predicted_frame.get("frame_id", -1)
for bbox in predicted_frame.get("boxes", []) or []:
try:
raw_cls = bbox.get("cls_id")
try:
object_id = int(raw_cls)
except (TypeError, ValueError):
object_id = None
looked_up = (
OBJECT_ID_LOOKUP.get(object_id) if object_id is not None else None
)
object_type: ObjectOfInterest
object_colour: ShirtColor = ShirtColor.OTHER
if looked_up is None:
object_type = ObjectOfInterest.PLAYER
elif isinstance(looked_up, str):
team_str = looked_up.strip().lower().replace(" ", "")
object_type = ObjectOfInterest.PLAYER
if team_str == "team1":
object_colour = TEAM1_SHIRT_COLOUR
elif team_str == "team2":
object_colour = TEAM2_SHIRT_COLOUR
else:
object_colour = ShirtColor.OTHER
else:
object_type = looked_up
team_field = (
(bbox.get("team") or bbox.get("team_id") or "").strip().lower()
)
if team_field in {"1", "team1"}:
object_colour = TEAM1_SHIRT_COLOUR
elif team_field in {"2", "team2"}:
object_colour = TEAM2_SHIRT_COLOUR
else:
object_colour = ShirtColor.OTHER
bboxes.append(
BoundingBox(
bbox_2d=[
int(bbox["x1"]),
int(bbox["y1"]),
int(bbox["x2"]),
int(bbox["y2"]),
],
label=object_type,
cluster_id=object_colour,
)
)
except Exception as e:
logger.error(e)
continue
miner_annotations[frame_number] = {
"bboxes": bboxes,
"action": predicted_frame.get("action", None),
"keypoints": predicted_frame.get("keypoints", []),
}
return miner_annotations
def post_vlm_ranking(
payload: TVPredictInput,
miner_run: SVRunOutput,
challenge: SVChallenge,
pseudo_gt_annotations: list[PseudoGroundTruth],
frame_store: FrameStore,
) -> SVEvaluation:
score_breakdown = TotalScore()
settings = get_settings()
miner_annotations = parse_miner_prediction(miner_run=miner_run)
logger.info(payload.meta)
challenge_type = challenge.challenge_type
if challenge_type is None:
challenge_type = parse_challenge_type(payload.meta.get("challenge_type"))
if (
miner_run.success
and len(miner_annotations) == settings.SCOREVISION_VIDEO_MAX_FRAME_NUMBER
and challenge_type is not None
):
score_breakdown.keypoints.floor_markings_alignment = evaluate_keypoints(
frames=frame_store,
miner_predictions=miner_annotations,
challenge_type=challenge_type,
)
score_breakdown.objects.bbox_placement = compare_object_placement(
pseudo_gt=pseudo_gt_annotations, miner_predictions=miner_annotations
)
score_breakdown.objects.categorisation = compare_object_labels(
pseudo_gt=pseudo_gt_annotations, miner_predictions=miner_annotations
)
score_breakdown.objects.team = compare_team_labels(
pseudo_gt=pseudo_gt_annotations, miner_predictions=miner_annotations
)
score_breakdown.objects.enumeration = compare_object_counts(
pseudo_gt=pseudo_gt_annotations, miner_predictions=miner_annotations
)
score_breakdown.objects.tracking_stability = bbox_smoothness_per_type(
video_bboxes=[
miner_annotations[frame_num]["bboxes"]
for frame_num in sorted(miner_annotations.keys())
],
image_height=settings.SCOREVISION_IMAGE_HEIGHT,
image_width=settings.SCOREVISION_IMAGE_WIDTH,
)
score_breakdown.latency.inference = 1 / 2 ** (miner_run.latency_ms / 1000)
else:
logger.info(
f"Miner success={miner_run.success} frames={len(miner_annotations)} "
f"challenge_type={getattr(challenge_type, 'value', None)} (must not be None)."
)
breakdown_dict = score_breakdown.to_dict()
objects_dict = breakdown_dict.get("objects", {}) or {}
keypoints_dict = breakdown_dict.get("keypoints", {}) or {}
def _mean_defined(values) -> float:
nums = [v for v in values if isinstance(v, (int, float))]
return (sum(nums) / len(nums)) if nums else 0.0
objects_score = _mean_defined(objects_dict.values())
keypoints_score = _mean_defined(keypoints_dict.values())
final_score = 0.5 * objects_score + 0.5 * keypoints_score
details = {
"breakdown": breakdown_dict,
"group_scores": {
"objects": objects_score,
"keypoints": keypoints_score,
},
"challenge": {
"id_hash": challenge.challenge_id,
"api_task_id": challenge.api_task_id,
"type": getattr(challenge.challenge_type, "value", None),
},
"prompt": challenge.prompt,
}
logger.info(details)
return SVEvaluation(
acc_breakdown=breakdown_dict,
latency_ms=miner_run.latency_ms,
acc=final_score,
score=final_score,
details=details,
)
|