hou-yz/MVDet

Lightweight MVDet.

zhangky666 opened this issue · 3 comments

zhangky666 commented

Thank you for this article, I have a great reference! Now I want to apply it to the project, using the Jetson SUB DEVELOPER KIT. Its CPU and GPU have a combined memory of 16GB, which is obviously not compatible with MVDET's requirement for two graphics cards with more than 11GB of memory. So I would like to ask if there is a MVDET lightweight approach, or is there a lightweight approach to optimize it? Hope to hear from you soon.Best wishes.

hou-yz commented

thank you for your interest!

please check out our other repo MVDeTr and specify to use the conv architecture python main.py -d wildtrack --world_feat conv. this should return similar results to the vanilla MVDet and only takes one RTX 2080TI to run.

hope this helps.

best,
Yunzhong

zhangky666 commented

Thanks for your reply. When I type "python main.py -d wildtrack --world_feat conv" in the terminal, it returns ”main.py: error: unrecognized arguments: --world_feat conv“. Is there anything I missed?

usage: main.py [-h] [--reID] [--cls_thres CLS_THRES] [--alpha ALPHA] [--variant {default,img_proj,res_proj,no_joint_conv}] [--arch {vgg11,resnet18}]
[-d {wildtrack,multiviewx}] [-j NUM_WORKERS] [-b N] [--epochs N] [--lr LR] [--weight_decay WEIGHT_DECAY] [--momentum M] [--log_interval N]
[--resume RESUME] [--visualize] [--seed SEED]

hou-yz commented

please use the MVDeTr repo instead of this one. this line python main.py -d wildtrack --world_feat conv refers to the main.py file in that repo.