Revert "约束预测框互不重叠"（eba533a、9439cbb）
去掉强制非重叠：detect.py 不再给 DetectionValidator.postprocess 打补丁， 配置里移除 train.max_overlap 与 agnostic_nms，可视化脚本恢复为直接使用模型原始输出， engine/detection 里的 suppress_overlaps / filter_overlapping_outputs 一并移除。 约束会把 mAP50 从 0.555 压到 0.483（precision 0.536->0.612），代价是按置信度抑制时 高分错框会压掉本可命中的低分框。需要时用 git revert 这次提交即可取回。 按类别分目录的可视化产出结构（36fff98）保留。 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
Shaun Zhou
co-authored by
Claude Opus 5
Claude Opus 5
parent
36fff984b7
commit
154677032f
@@ -48,13 +48,6 @@ train:
|
||||
hsv_s: 0.3
|
||||
hsv_v: 0.3
|
||||
|
||||
# 真值里一处缺陷只有一个框、零重叠（实测 11211 张图、151 张多框图，重叠为 0），
|
||||
# 所以验证时也把预测约束成互不重叠：早停与 best.pt 的选择才和最终推理一致。
|
||||
# 负数=不约束。注意这不是 ultralytics 的参数，由 detect.py 自己实现。
|
||||
max_overlap: 0.0
|
||||
# 端到端头的 top-k 每个 anchor 只取一个类别，压掉同一处缺陷被两个类别各框一次
|
||||
agnostic_nms: true
|
||||
|
||||
plots: true # results.png / PR 曲线 / 混淆矩阵 / val_batch 预测图
|
||||
save_period: -1
|
||||
|
||||
|
||||
@@ -6,11 +6,8 @@
|
||||
--checkpoint runs/detect/film/weights/best.pt
|
||||
|
||||
数据是在 Label Studio 里修正过的标签，导出并划分 train/val 之后放在 data/ 下的数据集。
|
||||
框来自 Grad-CAM / SHAP 的弱监督标注再经人工修正。
|
||||
|
||||
**预测框约束为互不重叠**（`train.max_overlap`，默认 0）。真值里一处缺陷只有一个框、
|
||||
零重叠，模型却会输出重叠框，所以验证阶段补一层抑制——早停与 best.pt 的选择都据此进行，
|
||||
指标才和最终推理一致。设为负数可关掉。
|
||||
框来自 Grad-CAM / SHAP 的弱监督标注，不是人工标注——指标该按"与热力区域的一致性"
|
||||
来读，而不是当成真值上的检测精度。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
@@ -22,7 +19,6 @@ from pathlib import Path
|
||||
|
||||
import yaml
|
||||
|
||||
from engine.detection import filter_overlapping_outputs
|
||||
from utils import load_config, seed_everything, setup_logging
|
||||
from utils.device import DeviceNotAvailableError, describe_device, resolve_device
|
||||
|
||||
@@ -39,23 +35,6 @@ def parse_args() -> argparse.Namespace:
|
||||
return parser.parse_args()
|
||||
|
||||
|
||||
def install_overlap_suppression(max_overlap: float) -> None:
|
||||
"""给验证器的后处理包一层"框不重叠"约束。
|
||||
|
||||
只能用打补丁的方式：ultralytics 没有对应的开关——端到端模型（yolo26）的 NMS 早退，
|
||||
传入的 iou / agnostic_nms 都不参与 IoU 抑制（见 engine/detection.filter_overlapping_outputs
|
||||
的说明）。补在 postprocess 之后，训练期验证、早停、best.pt 的选择就都遵守这条约束。
|
||||
"""
|
||||
from ultralytics.models.yolo.detect import DetectionValidator
|
||||
|
||||
original = DetectionValidator.postprocess
|
||||
|
||||
def postprocess(self, preds):
|
||||
return filter_overlapping_outputs(original(self, preds), max_iou=max_overlap)
|
||||
|
||||
DetectionValidator.postprocess = postprocess
|
||||
|
||||
|
||||
def to_ultralytics_device(device) -> str | int:
|
||||
"""torch.device -> ultralytics 的 device 记法（cuda 用序号，其余用字符串）。"""
|
||||
if device.type == "cuda":
|
||||
@@ -100,18 +79,10 @@ def main() -> int:
|
||||
|
||||
train_cfg = dict(cfg.get("train", {}))
|
||||
swanlab_cfg = dict(train_cfg.pop("swanlab", {}))
|
||||
# 不是 ultralytics 的参数，得先摘出来，否则它会拒绝未知键
|
||||
max_overlap = float(train_cfg.pop("max_overlap", 0.0))
|
||||
|
||||
from ultralytics import YOLO
|
||||
from ultralytics import settings as ul_settings
|
||||
|
||||
if max_overlap >= 0:
|
||||
install_overlap_suppression(max_overlap)
|
||||
logger.info("验证时约束预测框互不重叠（最大 IoU %.2f）", max_overlap)
|
||||
else:
|
||||
logger.warning("未约束预测框重叠：指标里会计入真值中不可能出现的重叠框")
|
||||
|
||||
# ultralytics 默认把数据集根指到 ~/datasets 之类的全局目录；我们的 data.yaml 里写的是
|
||||
# 绝对路径，但它仍会拿 datasets_dir 去拼相对路径，统一指到项目根省得踩坑。
|
||||
ul_settings.update({"datasets_dir": str(Path.cwd())})
|
||||
|
||||
@@ -102,84 +102,6 @@ def match_boxes(
|
||||
)
|
||||
|
||||
|
||||
def suppress_overlaps(
|
||||
boxes: Sequence[Box], max_iou: float = 0.0, contained: bool = True
|
||||
) -> list[int]:
|
||||
"""贪心地挑出互不重叠的框，返回保留下来的下标（按输入顺序）。
|
||||
|
||||
按置信度从高到低考察，与任一已保留框重叠超过 `max_iou` 就丢弃。
|
||||
|
||||
为什么需要这一步：NMS 只在**同类**内、且按固定 IoU 阈值抑制，所以同一处缺陷被两个
|
||||
类别各框一次（跨类重复），或两个框 IoU 低于 NMS 阈值时，都会留下重叠框。而人工标注的
|
||||
真值里一处缺陷只有一个框，零重叠——推理侧不加这层约束，指标和图上都会出现真值里
|
||||
不可能有的形态。
|
||||
|
||||
Args:
|
||||
max_iou: 允许的最大 IoU。0 表示碰到就算重叠，完全不允许。
|
||||
contained: 是否额外丢弃被已保留框完全包住的框。整图大框套小框时 IoU 可能很小
|
||||
（面积悬殊），但那显然是同一处目标的重复表达。
|
||||
"""
|
||||
order = sorted(range(len(boxes)), key=lambda i: -(boxes[i].score or 0.0))
|
||||
kept: list[int] = []
|
||||
for index in order:
|
||||
box = boxes[index]
|
||||
if any(
|
||||
iou(box, boxes[other]) > max_iou or (contained and _inside(box, boxes[other]))
|
||||
for other in kept
|
||||
):
|
||||
continue
|
||||
kept.append(index)
|
||||
return sorted(kept)
|
||||
|
||||
|
||||
def _inside(inner: Box, outer: Box) -> bool:
|
||||
"""inner 是否完全落在 outer 里（含边界）。"""
|
||||
return (
|
||||
outer.x1 <= inner.x1 and outer.y1 <= inner.y1
|
||||
and outer.x2 >= inner.x2 and outer.y2 >= inner.y2
|
||||
)
|
||||
|
||||
|
||||
def filter_overlapping_outputs(
|
||||
outputs: list[dict], max_iou: float = 0.0, contained: bool = True
|
||||
) -> list[dict]:
|
||||
"""给 ultralytics 验证器 postprocess 的输出补上"框不重叠"约束。
|
||||
|
||||
每个元素是一张图的结果：`bboxes`（xyxy）、`conf`、`cls` 等张量。坐标是不是归一化的
|
||||
无所谓，IoU 与尺度无关。
|
||||
|
||||
为什么要在这里补：yolo26 这类**端到端模型不走 IoU 抑制**。ultralytics 的
|
||||
`non_max_suppression` 对它们早退（`utils/nms.py` 里 `prediction.shape[-1] == 6 or end2end`
|
||||
的分支），只按置信度与 `max_det` 截断，传入的 `iou` 和 `agnostic` 都不生效；
|
||||
`agnostic_nms` 只作用于检测头的 top-k（每个 anchor 取一个类别），压不掉不同 anchor
|
||||
产生的重叠框。不补这一步，训练期的验证指标就和最终推理对不上。
|
||||
"""
|
||||
filtered = []
|
||||
for output in outputs:
|
||||
boxes = [
|
||||
Box(int(class_id), *(float(v) for v in xyxy), float(score))
|
||||
for xyxy, score, class_id in zip(
|
||||
output["bboxes"].tolist(), output["conf"].tolist(), output["cls"].tolist(),
|
||||
strict=True,
|
||||
)
|
||||
]
|
||||
kept = suppress_overlaps(boxes, max_iou=max_iou, contained=contained)
|
||||
if len(kept) == len(boxes):
|
||||
filtered.append(output)
|
||||
continue
|
||||
# 张量用列表下标筛选；空结果不能用 [] 索引，改用零长切片
|
||||
filtered.append({
|
||||
key: (value[kept] if kept else value[:0]) if _is_per_box(value, len(boxes)) else value
|
||||
for key, value in output.items()
|
||||
})
|
||||
return filtered
|
||||
|
||||
|
||||
def _is_per_box(value, count: int) -> bool:
|
||||
"""这个张量是不是"每个框一行"——只有这种才跟着筛选。"""
|
||||
return hasattr(value, "shape") and len(getattr(value, "shape", ())) >= 1 and len(value) == count
|
||||
|
||||
|
||||
def draw_boxes(image, ground_truth: Sequence[Box], predictions: Sequence[Box],
|
||||
classes: Sequence[str], match: Match, width: int = 2):
|
||||
"""把真值与预测画在图上（就地修改传入的 PIL 图像并返回）。
|
||||
|
||||
@@ -25,10 +25,6 @@ train / val 不分目录，但 `index.csv` 里有 `split` 列——训练集的
|
||||
同样地，"命中 / 漏检 / 误检"记在 `index.csv` 的 `outcome` 列里，按它排序筛选即可。
|
||||
|
||||
`--only errors` 只输出有漏检或误检的图，扫起来快得多。
|
||||
|
||||
**预测框默认约束为互不重叠**（`--max-overlap 0`）。人工标注的真值里一处缺陷只有一个框、
|
||||
零重叠，而 NMS 只在同类内按固定阈值抑制，留得下跨类重复和 IoU 低于阈值的重叠框。
|
||||
`--max-overlap -1` 可关掉这层约束，看模型的原始输出。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
@@ -46,13 +42,7 @@ sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
|
||||
import yaml # noqa: E402
|
||||
|
||||
from datasets.yolo import IMAGE_SUFFIXES # noqa: E402
|
||||
from engine.detection import ( # noqa: E402
|
||||
Box,
|
||||
draw_boxes,
|
||||
match_boxes,
|
||||
parse_label,
|
||||
suppress_overlaps,
|
||||
)
|
||||
from engine.detection import Box, draw_boxes, match_boxes, parse_label # noqa: E402
|
||||
from utils import setup_logging # noqa: E402
|
||||
from utils.device import DeviceNotAvailableError, describe_device, resolve_device # noqa: E402
|
||||
|
||||
@@ -133,16 +123,6 @@ def main() -> int:
|
||||
parser.add_argument("--imgsz", type=int, default=640)
|
||||
parser.add_argument("--conf", type=float, default=0.25, help="置信度阈值")
|
||||
parser.add_argument("--iou-threshold", type=float, default=0.5, help="判定命中的 IoU")
|
||||
parser.add_argument(
|
||||
"--max-overlap", type=float, default=0.0,
|
||||
help="预测框之间允许的最大 IoU，超出的丢低分框。0=完全不允许重叠（默认，与真值一致），"
|
||||
"-1=不约束，看模型原始输出",
|
||||
)
|
||||
parser.add_argument("--nms-iou", type=float, default=0.7, help="NMS 的 IoU 阈值")
|
||||
parser.add_argument(
|
||||
"--no-agnostic-nms", action="store_true",
|
||||
help="关掉类别无关 NMS。默认开启，用来压掉同一处缺陷被两个类别各框一次的重复",
|
||||
)
|
||||
parser.add_argument("--batch", type=int, default=16)
|
||||
parser.add_argument(
|
||||
"--only", choices=("all", "errors"), default="all",
|
||||
@@ -170,12 +150,7 @@ def main() -> int:
|
||||
"数据集: %s（%d 类，%s）", args.data, len(classes),
|
||||
"，".join(f"{name} {len(images)} 张" for name, images in splits.items()),
|
||||
)
|
||||
logger.info(
|
||||
"conf=%.2f 命中 IoU=%.2f imgsz=%d NMS IoU=%.2f 类别无关 NMS=%s 框间最大重叠=%s",
|
||||
args.conf, args.iou_threshold, args.imgsz, args.nms_iou,
|
||||
not args.no_agnostic_nms,
|
||||
"不约束" if args.max_overlap < 0 else f"{args.max_overlap:.2f}",
|
||||
)
|
||||
logger.info("conf=%.2f IoU=%.2f imgsz=%d", args.conf, args.iou_threshold, args.imgsz)
|
||||
|
||||
if out_dir.exists() and any(out_dir.iterdir()):
|
||||
if not args.overwrite:
|
||||
@@ -196,7 +171,6 @@ def main() -> int:
|
||||
scores: dict[str, list[float]] = defaultdict(list)
|
||||
rows = []
|
||||
processed = 0
|
||||
suppressed = 0
|
||||
|
||||
def predict_in_batches(images):
|
||||
"""自己按批切分：ultralytics 的 predict 不理会 batch 参数，
|
||||
@@ -205,8 +179,7 @@ def main() -> int:
|
||||
chunk = images[start:start + args.batch]
|
||||
results = model.predict(
|
||||
source=[str(p) for p in chunk], imgsz=args.imgsz, conf=args.conf,
|
||||
device=device_arg, verbose=False, iou=args.nms_iou,
|
||||
agnostic_nms=not args.no_agnostic_nms,
|
||||
device=device_arg, verbose=False,
|
||||
)
|
||||
yield from zip(chunk, results, strict=True)
|
||||
|
||||
@@ -219,11 +192,6 @@ def main() -> int:
|
||||
result.boxes.conf.tolist(), strict=True,
|
||||
)
|
||||
]
|
||||
if args.max_overlap >= 0:
|
||||
kept = suppress_overlaps(predictions, max_iou=args.max_overlap)
|
||||
suppressed += len(predictions) - len(kept)
|
||||
predictions = [predictions[i] for i in kept]
|
||||
|
||||
label_file = label_path_for(image_path)
|
||||
truth = parse_label(label_file.read_text(encoding="utf-8")) if label_file.exists() else []
|
||||
match = match_boxes(predictions, truth, args.iou_threshold)
|
||||
@@ -273,10 +241,6 @@ def main() -> int:
|
||||
"conf": args.conf,
|
||||
"iou_threshold": args.iou_threshold,
|
||||
"imgsz": args.imgsz,
|
||||
"nms_iou": args.nms_iou,
|
||||
"agnostic_nms": not args.no_agnostic_nms,
|
||||
"max_overlap": args.max_overlap,
|
||||
"suppressed_overlapping_boxes": suppressed,
|
||||
"outcomes": {split: dict(counter) for split, counter in outcomes.items()},
|
||||
"per_class": {
|
||||
split: {
|
||||
@@ -293,8 +257,6 @@ def main() -> int:
|
||||
json.dumps(summary, ensure_ascii=False, indent=2), encoding="utf-8"
|
||||
)
|
||||
|
||||
if suppressed:
|
||||
logger.info("因与更高分的框重叠而丢弃: %d 个", suppressed)
|
||||
for split, counter in outcomes.items():
|
||||
shown = sum(counter[k] for k in ("missed", "extra"))
|
||||
logger.info(
|
||||
|
||||
@@ -10,12 +10,10 @@ from engine.detection import ( # noqa: E402
|
||||
COLOR_UNMATCHED,
|
||||
Box,
|
||||
draw_boxes,
|
||||
filter_overlapping_outputs,
|
||||
from_yolo,
|
||||
iou,
|
||||
match_boxes,
|
||||
parse_label,
|
||||
suppress_overlaps,
|
||||
)
|
||||
|
||||
|
||||
@@ -131,117 +129,3 @@ def test_draw_handles_boxes_at_the_border():
|
||||
prediction = Box(0, 0.0, 0.0, 0.5, 0.05, 0.9)
|
||||
draw_boxes(image, [], [prediction], ["blackdot"], match_boxes([prediction], []))
|
||||
assert COLOR_UNMATCHED in _colors(image)
|
||||
|
||||
|
||||
# ---------------- 不重叠约束 ----------------
|
||||
|
||||
|
||||
def test_suppress_keeps_the_highest_score_of_an_overlapping_pair():
|
||||
"""同一处缺陷被框两次时，留置信度高的那个。"""
|
||||
strong = Box(0, 0.3, 0.3, 0.7, 0.7, 0.9)
|
||||
weak = Box(1, 0.35, 0.35, 0.75, 0.75, 0.4)
|
||||
assert suppress_overlaps([weak, strong]) == [1] # 下标 1 = strong
|
||||
|
||||
|
||||
def test_suppress_is_cross_class():
|
||||
"""NMS 只在同类内抑制；跨类重复要靠这一步。"""
|
||||
a = Box(0, 0.3, 0.3, 0.7, 0.7, 0.9)
|
||||
b = Box(5, 0.31, 0.31, 0.69, 0.69, 0.8)
|
||||
assert suppress_overlaps([a, b]) == [0]
|
||||
|
||||
|
||||
def test_suppress_keeps_disjoint_boxes():
|
||||
"""真正分开的多个缺陷不能被误伤。"""
|
||||
boxes = [
|
||||
Box(0, 0.0, 0.0, 0.2, 0.2, 0.9),
|
||||
Box(0, 0.5, 0.5, 0.7, 0.7, 0.8),
|
||||
Box(1, 0.8, 0.1, 0.95, 0.3, 0.7),
|
||||
]
|
||||
assert suppress_overlaps(boxes) == [0, 1, 2]
|
||||
|
||||
|
||||
def test_touching_edges_are_not_overlap():
|
||||
"""边缘相接、面积不相交，IoU=0，应当都保留。"""
|
||||
left = Box(0, 0.0, 0.0, 0.5, 1.0, 0.9)
|
||||
right = Box(0, 0.5, 0.0, 1.0, 1.0, 0.8)
|
||||
assert suppress_overlaps([left, right]) == [0, 1]
|
||||
|
||||
|
||||
def test_contained_box_is_dropped_even_when_iou_is_below_the_threshold():
|
||||
"""整图大框套一个小框：面积悬殊导致 IoU 很小，但显然是同一处目标的重复表达。
|
||||
|
||||
这正是 `contained` 的用武之地——只在允许轻微重叠（max_iou>0）时才有区别，
|
||||
因为 max_iou=0 时任何包含都已经因 IoU>0 被抑制了。
|
||||
"""
|
||||
big = Box(0, 0.0, 0.0, 1.0, 1.0, 0.9)
|
||||
small = Box(1, 0.45, 0.45, 0.55, 0.55, 0.8)
|
||||
assert iou(big, small) < 0.02
|
||||
|
||||
assert suppress_overlaps([big, small], max_iou=0.1) == [0]
|
||||
assert suppress_overlaps([big, small], max_iou=0.1, contained=False) == [0, 1]
|
||||
|
||||
|
||||
def test_max_iou_allows_slight_overlap():
|
||||
a = Box(0, 0.0, 0.0, 0.5, 0.5, 0.9)
|
||||
b = Box(0, 0.4, 0.4, 0.9, 0.9, 0.8) # IoU 约 0.02
|
||||
assert suppress_overlaps([a, b]) == [0]
|
||||
assert suppress_overlaps([a, b], max_iou=0.1) == [0, 1]
|
||||
|
||||
|
||||
def test_suppress_handles_empty_and_single():
|
||||
assert suppress_overlaps([]) == []
|
||||
assert suppress_overlaps([Box(0, 0.1, 0.1, 0.2, 0.2, 0.5)]) == [0]
|
||||
|
||||
|
||||
# ---------------- 验证器输出的过滤 ----------------
|
||||
|
||||
|
||||
def test_filter_overlapping_outputs_drops_the_lower_score():
|
||||
"""ultralytics 验证器的输出是张量字典，筛选要对每个"每框一行"的张量同步生效。"""
|
||||
torch = pytest.importorskip("torch")
|
||||
|
||||
output = {
|
||||
"bboxes": torch.tensor([[0.0, 0.0, 10.0, 10.0], [1.0, 1.0, 11.0, 11.0],
|
||||
[50.0, 50.0, 60.0, 60.0]]),
|
||||
"conf": torch.tensor([0.9, 0.5, 0.8]),
|
||||
"cls": torch.tensor([0.0, 1.0, 2.0]),
|
||||
"extra": torch.zeros((3, 0)),
|
||||
}
|
||||
filtered = filter_overlapping_outputs([output])[0]
|
||||
|
||||
assert filtered["conf"].tolist() == pytest.approx([0.9, 0.8]) # 中间那个与第一个重叠
|
||||
assert filtered["cls"].tolist() == [0.0, 2.0]
|
||||
assert filtered["bboxes"].shape == (2, 4)
|
||||
assert filtered["extra"].shape == (2, 0)
|
||||
|
||||
|
||||
def test_filter_keeps_the_object_untouched_when_nothing_overlaps():
|
||||
torch = pytest.importorskip("torch")
|
||||
|
||||
output = {
|
||||
"bboxes": torch.tensor([[0.0, 0.0, 1.0, 1.0], [5.0, 5.0, 6.0, 6.0]]),
|
||||
"conf": torch.tensor([0.9, 0.8]),
|
||||
"cls": torch.tensor([0.0, 0.0]),
|
||||
}
|
||||
filtered = filter_overlapping_outputs([output])
|
||||
assert filtered[0] is output # 没有变化时原样返回，不做无谓的拷贝
|
||||
|
||||
|
||||
def test_filter_handles_images_with_no_detection():
|
||||
torch = pytest.importorskip("torch")
|
||||
|
||||
empty = {"bboxes": torch.zeros((0, 4)), "conf": torch.zeros(0), "cls": torch.zeros(0)}
|
||||
assert filter_overlapping_outputs([empty])[0]["conf"].numel() == 0
|
||||
|
||||
|
||||
def test_filter_can_empty_an_image_entirely():
|
||||
"""全部互相重叠时只留一个；这里构造两个几乎重合的框。"""
|
||||
torch = pytest.importorskip("torch")
|
||||
|
||||
output = {
|
||||
"bboxes": torch.tensor([[0.0, 0.0, 10.0, 10.0], [0.1, 0.1, 10.1, 10.1]]),
|
||||
"conf": torch.tensor([0.3, 0.9]),
|
||||
"cls": torch.tensor([1.0, 1.0]),
|
||||
}
|
||||
filtered = filter_overlapping_outputs([output])[0]
|
||||
assert filtered["conf"].tolist() == pytest.approx([0.9])
|
||||
|
||||
Reference in New Issue
Block a user