CarlHuang

Kalman filter fails in low speed scenario

4

Hi , in you papars, says Kalman filter fails in autonomous driving scenario. So Which motion Filter should using ??

change NMS to soft_nms

1

Whether change nms to soft_nms for dence detection environment ? Because we encountered many missed detection using defeat YOLOX configuration in Dense crowd detection environment.

HI Zhongdao, I want to generate feature embedding from intance mask by UniTrack Module(or pretrained res50 model)? Whether Unitrack framework can support ? Thanks, Carl

Hi Zhongdao , Do you have a quantitative reid indicator(mAP) for each model on different data sets(such as MOT)??

download pretrain model issue

2

Hi, Zhongdao, I can not download pretrain model from model zoo links,:https://github.com/Zhongdao/UniTrack/blob/54347ba1bdba0903b241e00de2b5d0dc3c1a3d14/docs/MODELZOO.md (ImageNet classification Imagenet 50 & SimCLR v2 )

reid

1

Why your retrain Reid module base on MOT17 data have a very high accuracy(mAP > 80) from your reid_log(https://download.openmmlab.com/mmtracking/fp16/reid_r50_fp16_8x32_6e_mot17_20210731_033055.log.json) but we reproduce it ,the accuracy always ~60 (mAP)

reid accuracy

4

Why retrain reid config alway get very low accuracy(mAP:50), From your log & pretrain model, it can achieve about 96% mAP.

Hope framework can support VIS (mask2former) ASAP

1

RT, Hope framework can support VIS (mask2former) ASAP which is new SOTA in VIS direction.

Hi, How to split your Unicorn framework as Detection, Apperance Model(embeding), association model? So we can easy take replace of different module.

Apperance Model performance very bad in night scenario ?

1

Hi Author, I found Apperance Model can work in many scenario. but in night ,the picture quality or light have big impact on Apperance model peformance . Why it is...