DetVisGUI

Introuction

This is a lightweight GUI for visualizing the results of deep learning object detection (such as mmdetection). It could display detection results with different threshold dynamically, and would be convenient for verifying detection results and groundtruth.

DetVisGUI supports Pascal VOC and COCO formats.

Video with text description : https://www.youtube.com/watch?v=4imQyECTik0

Dependencies

numpy
matplotlib
cv2
cocoapi

Code

Clone this repository.

git clone https://github.com/Chien-Hung/DetVisGUI.git
cd DetVisGUI

Demo

I sample a small part of COCO and VOC2007 dataset, running mmdetection for getting detection result(*.pkl) and use these files for demo. There is not any information about image and annotation in the detection result(*.pkl), so I link the image, annotation, [image list text file(VOC)] and detection result(*.pkl) by arguments.

python DetVisGUI.py --format ${DATASET_FORMAT} --img_root ${IMAGE_ROOT} --anno_root ${ANNOTATION_ROOT} --txt ${IMAGE_LIST} --det_file ${DETECTION_FILE} [--output ${SAVE_DIRECTORY}] [--no_gt]

Arguments:

DATASET_FORMAT: select dataset format, COCO or VOC.
IMAGE_ROOT: The image directory.
ANNOTATION_ROOT: The annotation directory (VOC) / annotation json file (COCO).
IMAGE_LIST: The path of image list txt file. This argument is only for VOC format.
DETECTION_FILE: The detection output file (*.pkl).

Optional Arguments:

SAVE_DIRECTORY: The directory for saving display images.
--no_gt: If there are no annotations according to display images (test images), add --no_gt.

Display COCO validation results:

$ python DetVisGUI.py --format COCO \
      --img_root data/COCO/val2017 \
      --anno_root data/COCO/instances_val2017.json \
      --det_file results/mask_rcnn_r50_fpn_1x/val_results.pkl

Display COCO test results (no groundtruth):

$ python DetVisGUI.py --format COCO \
      --img_root data/COCO/test2017 \
      --anno_root data/COCO/image_info_test-dev2017.json \
      --det_file results/mask_rcnn_r50_fpn_1x/test_results.pkl \
      --no_gt

Display Pascal VOC training results:

$ python DetVisGUI.py --format VOC \
      --img_root data/VOCdevkit/VOC2007/JPEGImages \
      --anno_root data/VOCdevkit/VOC2007/Annotations \
      --txt data/VOCdevkit/VOC2007/ImageSets/Main/train.txt \
      --det_file results/ssd512_voc/train_results.pkl

Display Pascal VOC test results (no groundtruth):

$ python DetVisGUI.py --format VOC \
      --img_root data/VOCdevkit/VOC2007/JPEGImages \
      --anno_root data/VOCdevkit/VOC2007/Annotations \
      --txt data/VOCdevkit/VOC2007/ImageSets/Main/test.txt \
      --det_file results/ssd512_voc/test_results.pkl \
      --no_gt

Hotkeys

KEY	ACTION
↑ , ↓	change image.
← , →	change score threshold.
ctrl + ← , →	change IoU threshold.
s	save displayed image in output folder.
q	colse this GUI.

zhjpqq / DetVisGUI