Using the hrnet_w32_h36m_256x256 model to perform inference on custom data but cannot get correct results

[Lanrzip]

When I use the hrnet_w32_h36m_256x256 model and the top_down_img_demo_with_mmdet.py file to inference the h36m data set, the model can give me the correct result, but when I use the model to inference the custom image, the output keypoints score is very low (such as 1.27726495 e-02). May I ask why?
The specific model and code I use are as follows:

python mmpose/demo/top_down_img_demo_with_mmdet.py \
    mmdetection/myconfigs/fcos_r50_caffe_fpn_gn-head_mstrain_640-800_2x_coco_copy.py \
    mmdetection/work_dirs/fcos_r50_caffe_fpn_gn-head_mstrain_640-800_2x_coco_copy/latest.pth \
    mmpose/configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/h36m/hrnet_w32_h36m_256x256.py \
    mmpose/pth_files/hrnet_w32_h36m_256x256-d3206675_20210621.pth \
    --img-root standard_pose/ \
    --img t4.png \
    --out-img-root standard_pose