JfortissJ / Deformable_detr_support

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

This code is for deformable-detr using support image

The file tree should be:

-work/

​	-deformable_detr_support/

-data/

​	-cityscape/

Preparation

dataset and pretrained weights

The custome cityscape dataset is in the mega cloud, user should download from

https://mega.nz/file/wXcTWQoR#ANd1kLJe1hh2A0LwHwUHZAfpAGVmyYWr4GHb0UiMd6I

There are two checkpoints for visualization, one for heat map and the other for detection results, download link:

For bounding box generation:

https://mega.nz/file/hDlhjCLR#OIlmslbyLkLCqmuYC4ZXfdpZSdByBfQX-CplUrA2sdA

For heatmap generation:

https://mega.nz/file/Mf0B1YoR#1Y36nWxgJTl5mZq-npiOOg9qONgDf6XCOclosgnd37o

After downloading the dataset, please create two soft links ./car_pool -> data/cityscape/car and ./person_pool -> data/cityscape/person in -work/deformable_detr_support.

model

After placing the model file in the correct location, please install the MultiScaleDeformableAttention by executing:

cd work/deformable_detr_support
./models/opt/make.sh

You should modify the checkpoint path in each .sh file before executing it.

Running

Visualizing detection results:

./generate_pred_bbox.sh

visualizing heatmap with reference points:

./generate_car_heat_map.sh

You could find the results in the output dictionary which is specified in each .sh file.

About


Languages

Language:Python 81.0%Language:Cuda 16.0%Language:C++ 1.6%Language:Shell 1.4%