Detectron2加载PointRend模型时出现pixel_mean/std未使用警告且检测终止

问题场景 #

运行基于Detectron2的短视频移动物体检测代码时，突然出现以下警告且检测终止：

D:\detectron2>python motion_detection_2.py
[02/27 12:39:29 d2.checkpoint.detection_checkpoint]: [DetectionCheckpointer] Loading from detectron2://PointRend/InstanceSegmentation/pointrend_rcnn_X_101_32x8d_FPN_3x_coco/28119989/model_final_ba17b9.pkl ...
WARNING [02/27 12:39:30 d2.projects.point_rend.mask_head]: Weight format of PointRend models have changed! Applying automatic conversion now ...
The checkpoint state_dict contains keys that are not used by the model:
  pixel_mean
  pixel_std

确认detection_checkpoint.py第129行处理pixel_mean和pixel_std的代码已执行，但问题仍存在。

解决方案 #

1. 手动过滤不需要的权重键 #

加载模型前，手动从checkpoint的state_dict中移除这两个冗余键，再加载至模型：

from detectron2.checkpoint import DetectionCheckpointer
from detectron2.config import get_cfg
from detectron2.projects import point_rend
from detectron2.modeling import build_model

# 初始化配置
cfg = get_cfg()
point_rend.add_pointrend_config(cfg)
# 补充你的模型配置（如数据集、推理参数等）
cfg.MODEL.WEIGHTS = "detectron2://PointRend/InstanceSegmentation/pointrend_rcnn_X_101_32x8d_FPN_3x_coco/28119989/model_final_ba17b9.pkl"

# 构建模型
model = build_model(cfg)
checkpointer = DetectionCheckpointer(model)

# 加载checkpoint并移除冗余键
checkpoint = checkpointer._load_file(cfg.MODEL.WEIGHTS)
if "model" in checkpoint:
    for key in ["pixel_mean", "pixel_std"]:
        checkpoint["model"].pop(key, None)

# 加载处理后的权重
checkpointer._load_model(checkpoint)

2. 显式配置预处理参数 #

在配置中直接指定pixel_mean和pixel_std，让模型从配置获取参数而非checkpoint：

# 在配置初始化后添加以下内容
cfg.MODEL.PIXEL_MEAN = [103.530, 116.280, 123.675]  # COCO数据集默认值
cfg.MODEL.PIXEL_STD = [57.375, 57.120, 58.395]

这样模型会优先使用配置中的参数，忽略checkpoint里的对应键。

3. 临时禁用警告（应急方案） #

如果上述方法无效，可临时禁用该模块的警告输出：

import logging
# 将检测检查点模块的日志级别设为ERROR，屏蔽警告
logging.getLogger("detectron2.checkpoint.detection_checkpoint").setLevel(logging.ERROR)

内容的提问来源于stack exchange，提问作者dkloska