Skip to content

PhyllisH/DVDET

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

111 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

DVDET

paper

This repository contains the official PyTorch implementation of

Aerial Monocular 3D Object Detection
Yue Hu, Shaoheng Fang, Weidi Xie, Siheng Chen
Presented at [RAL 2022]

[Dataset] can be downloaded at https://pan.baidu.com/s/1ZT9z4B5hvwJVFqwdEftkPQ?pwd=pdh3#list/path=%2F

Details

Args

exp_id: the path to save the models and logs

batch_size: the overall batch size

master_batch: the batch size of the master gpu (which maybe slightly smaller than the average batch size)

num_agents: the agent amount of a single sample

lr: the learning rate

gpus: the visible gpus; 0,1,2,3

num_epochs: the overall epoches

message_mode: NO_MESSAGE; this arg may be used in the collaborative setting

uav_height: the altitude of the drone used to colllect dataset, used to chose dataset; could be 40/60/80

map_scale: the default value is 1.0, which means the resolution of the BEV feature map is 0.25m/pixel

trans_layer: the layer of the feature map where collaboration operated; -2 means no collaboration

coord: the coordinate of object detection; Global means the BEV coordinate, Local means the image/camera coordinate; Joint means both BEV and Image coordinate

warp_mode: the method to transform feature/image to the BEV coordinate; HW means hard warping which transforms based on the projection matrix; DW means deformable warping which tranforms based on the residual of the projection matrix and the learnable deformable offsets; DADW means a distance-aware deformable warping which considers the geometric prior: the coordinates of the pixels; since the offset of the near pixels should be smaller than the far-away ones

depth_mode: the BEV feature map could be generated considering all the possible altitudes or only the ground plane whose altitude equals to zerop; the corresponding modes are Weighted and Unique

polygon: the object representation could be rotated rectangle or axis-aligned rectangle

real: if set true, the model would be trained on the real dataset, otherwise the virtual dataset

Train

CUDA_VISIBLE_DEVICES=GPU_ID python main.py multiagent_det --exp_id EXP_DIR --batch_size=BATCH_SIZE --master_batch=MASTER_BATCH_SIZE --num_agents=NUM_AGENTS --lr=LR --gpus GPU_ID --num_epochs EPOCHS --message_mode=NO_MESSAGE --uav_height=40 --map_scale=1.0 --trans_layer -2 --coord=Global/Local/Joint --feat_mode=FEAT_MODE --warp_mode=DW/HW/DADW --depth_mode=Weighted/Unique --polygon

Inference

CUDA_VISIBLE_DEVICES=GPU_ID python multiagent_test.py multiagent_det --exp_id EXP_DIR --load_model MODEL_DIR --gpus GPU_ID  --message_mode=NO_MESSAGE --uav_height=40 --map_scale=1.0 --trans_layer -2 --coord=Global/Local/Joint --feat_mode=FEAT_MODE  --warp_mode=DW/HW/DADW --depth_mode=Weighted/Unique --polygon

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors