PaddlePaddle / PaddleDetection

Object Detection toolkit based on PaddlePaddle. It supports object detection, instance segmentation, multiple object tracking and real-time multi-person keypoint detection.
Apache License 2.0
12.77k stars 2.88k forks source link

MTMCT对AIC21-demo中的车辆跟踪可视化结果存在过多boundingbox和频繁的IDS,该调整那些参数。 #6536

Open wolo-wolo opened 2 years ago

wolo-wolo commented 2 years ago

问题确认 Search before asking

请提出你的问题 Please ask your question

通过如下命令成功运行对两个demo视频的mtmct:

python deploy/pptracking/python/mot_sde_infer.py --model_dir=output_inference/ppyolov2_r50vd_dcn_365e_aic21mtmct_vehicle/ --reid_model_dir=output_inference/deepsort_pplcnet_vehicle/ --mtmct_dir=mtmct-demo --mtmct_cfg=deploy/pptracking/python/mtmct_cfg.yml --device=GPU --tracker_config=deploy/pptracking/python/tracker_config.yml --scaled=True --save_mot_txts --save_images --cpu_threads -1

输出的跟踪结果: 图片

我想为每个车辆仅展示一个boundingbox和ID,该调整那些参数,我尝试过修改如下超参数: output_inference/ppyolov2_r50vd_dcn_365e_aic21mtmct_vehicle/infer_cfg.yml中的draw_threshold; output_inference/deepsort_pplcnet_vehicle/infer_cfg.ymldeploy/pptracking/python/tracker_config.yml 中的draw_threshold,matching_threshold,和max_iou_distance。 均未在output/mtmct_vis得到满意的可视化效果。@nemonameless

nemonameless commented 2 years ago

以前的 ppyolov2车辆模型可能训的不够好。建议 model_dir 换成 https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_l_36e_ppvehicle.zip 试试,其实就是 https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppvehicle 这里的车辆模型导出后的,新训的模型精度更高,数据集场景覆盖的更全面。

wolo-wolo commented 2 years ago

以前的 ppyolov2车辆模型可能训的不够好。建议 model_dir 换成 https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_l_36e_ppvehicle.zip 试试,其实就是 https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppvehicle 这里的车辆模型导出后的,新训的模型精度更高,数据集场景覆盖的更全面。

非常感谢您的及时回复,我替换了该检测模型: 图片

图片

除了过多boundingbox和频繁的ID,似乎还表现出更糟的错误检测

wolo-wolo commented 2 years ago

@nemonameless @ZeyuChen @QingshuChen @sjtubinlong