Add vis2yolo utils, disp nb

adv_patch_gen/conv_visdrone_2_yolo/README.md

@@ -0,0 +1,84 @@
+# Adversarial Patches against UAE Object Detection
+
+## Setup
+
+Tested with python 3.8
+
+```shell
+python -m venv venv
+source venv/bin/activate
+pip install tqdm==4.65.0
+pip install imagesize==1.4.1
+pip install opencv-python==4.7.0.72
+```
+
+## VisDrone Dataset Format
+
+Dataset can be downloaded from <https://github.com/VisDrone/VisDrone-Dataset>.
+
+Annotations for the detections are follows:
+
+`<bbox_left>, <bbox_top>, <bbox_width>, <bbox_height>, <score>, <object_category>, <truncation>, <occlusion>`
+
+  - <bbox_left> The x coordinate of the top-left corner of the predicted bounding box
+  - <bbox_top> The y coordinate of the top-left corner of the predicted object bounding box
+  - <bbox_width> The width in pixels of the predicted object bounding box
+  - <bbox_height> The height in pixels of the predicted object bounding box
+  - <score> The score in the DETECTION file indicates the confidence of the predicted bounding box enclosing an object instance. The score in GROUNDTRUTH file is set to 1 or 0. 1 indicates the bounding box is considered in evaluation, while 0 indicates the bounding box will be ignored.
+  - <object_category> The object category indicates the type of annotated object, (i.e., ignored regions(0), pedestrian(1), people(2), bicycle(3), car(4), van(5), truck(6), tricycle(7), awning-tricycle(8), bus(9), motor(10), others(11))
+  - <truncation> The score in the DETECTION result file should be set to the constant -1.The score in the GROUNDTRUTH file indicates the degree of object parts appears outside a frame (i.e., no truncation = 0 (truncation ratio 0%), and partial truncation = 1 (truncation ratio 1% ~ 50%)).
+  - <occlusion> The score in the DETECTION file should be set to the constant -1. The score in the GROUNDTRUTH file indicates the fraction of objects being occluded (i.e., no occlusion = 0 (occlusion ratio 0%), partial occlusion = 1 (occlusion ratio 1% ~ 50%), and heavy occlusion = 2 (occlusion ratio 50% ~ 100%)).
+
+## Download VisDrone Dataset
+
+Download Dataset from <https://github.com/VisDrone/VisDrone-Dataset> and unzip and place under top level directory `data`.
+
+Alternatively, use gdown to download the zip files from the command line.
+
+```shell
+mkdir data
+cd data
+pip install gdown
+# object detection train subset
+gdown 1a2oHjcEcwXP8oUF95qiwrqzACb2YlUhn
+# object detection val subset
+gdown 1bxK5zgLn0_L8x276eKkuYA_FzwCIjb59
+# object detection test subset
+gdown 1PFdW_VFSCfZ_sTSZAGjQdifF_Xd5mf0V
+
+# unzip all data
+unzip VisDrone2019-DET-test-dev.zip
+unzip VisDrone2019-DET-train.zip
+unzip VisDrone2019-DET-val.zip
+
+# remove all zip files
+rm *.zip
+```
+
+## Visualize Images
+
+Running annotation visualizations for VisDrone or YOLO format.
+
+```shell
+python disp_visdrone.py -a ANNOTS_DIR -i IMAGES_DIR
+python disp_yolo.py -a ANNOTS_DIR -i IMAGES_DIR
+```
+
+## Convert VisDrone to YOLO annotation format
+
+Note: The classes to consider and any additional re-assingment of classes must be done with variables `CLASS_2_CONSIDER` and `CLASS_ID_REMAP` inside `conv_visdrone_2_yolo_fmt.py`.
+
+Can use optional params `low_dim_cutoff` and `low_area_cutoff` to cutoff bunding boxes that do not satisfy a minimum box dimenison or area percentage cutoff.
+
+```shell
+# example conversion of VisDrone train, val and test set annotations to YOLO format. Use -h for all options
+python conv_visdrone_2_yolo_fmt.py --sad data/VisDrone2019-DET-train/annotations --sid data/VisDrone2019-DET-train/images --td data/VisDrone2019-DET-train/labels
+python conv_visdrone_2_yolo_fmt.py --sad data/VisDrone2019-DET-val/annotations --sid data/VisDrone2019-DET-val/images --td data/VisDrone2019-DET-val/labels
+python conv_visdrone_2_yolo_fmt.py --sad data/VisDrone2019-DET-test-dev/annotations --sid data/VisDrone2019-DET-test-dev/images --td data/VisDrone2019-DET-test-dev/labels
+```
+
+Note: To convert to COCO annotation format refer to <https://github.com/SamSamhuns/ml_data_processing/tree/master/annotation_format_conv>
+
+## References
+
+- [VisDrone](https://github.com/VisDrone/VisDrone-Dataset)

adv_patch_gen/conv_visdrone_2_yolo/conv_visdrone_2_yolo_fmt.py

@@ -0,0 +1,116 @@
+"""
+Convert VisDrone annotation format to YOLO labels format
+Works for training of YOLOv5 and YOLOv7.
+YOLOv7 requires an additional txt file (Same name as the first parent directory) with paths to the images for the train, val & test splits
+"""
+import os
+import os.path as osp
+import glob
+import argparse
+from typing import Optional
+
+import tqdm
+import imagesize
+
+# VisDrone annot fmt
+# <bbox_left>, <bbox_top>, <bbox_width>, <bbox_height>, <score>, <object_category>, <truncation>, <occlusion>
+#
+# classes:
+# ignore(0), pedestrian(1), people(2), bicycle(3),
+# car(4), van(5), truck(6), tricycle(7),
+# awning-tricycle(8), bus(9), motor(10), others(11)
+
+# YOLO annot fmt
+# One row per object
+# Each row is class x_center y_center width height format.
+# Box coordinates must be in normalized xywh format (from 0 - 1).
+# If boxes are in pixels, divide x_center and width by image width, and y_center and height by image height.
+# Class numbers are zero-indexed (start from 0).
+
+
+CLASS_2_CONSIDER = {1, 2, 4, 5, 6, 9} # only get pedestrian, people, car, van, truck, bus classes from VisDrone
+CLASS_ID_REMAP = {4: 0, 5: 0, 6: 1, 9: 2, 1: 3, 2: 3} # optionally remap class ids, can be set to None
+IMG_EXT = {".jpg", ".png"}
+
+
+def get_parsed_args():
+ parser = argparse.ArgumentParser(
+ description="VisDrone to YOLO annot format")
+ parser.add_argument('--sad', '--source_annot_dir', type=str, dest="source_annot_dir", required=True,
+ help='VisDrone annotation source dir. Should contain annot txt files (default: %(default)s)')
+ parser.add_argument('--sid', '--source_image_dir', type=str, dest="source_image_dir", required=True,
+ help='VisDrone images source dir. Should contain image files (default: %(default)s)')
+ parser.add_argument('--td', '--target_annot_dir', type=str, dest="target_annot_dir", required=True,
+ help='YOLO annotation target dir. YOLO by default uses dirname labels (default: %(default)s)')
+ parser.add_argument('--dc', '--low_dim_cutoff', type=int, dest="low_dim_cutoff", default=None,
+ help='All bboxes with dims(w/h) < cutoff pixel are ignored i.e 400 (default: %(default)s)')
+ parser.add_argument('--ac', '--low_area_cutoff', type=float, dest="low_area_cutoff", default=None,
+ help='All bboxes with area perc < cutoff area perc are ignored i.e. 0.01 (default: %(default)s)')
+ args = parser.parse_args()
+ return args
+
+
+def conv_visdrone_2_yolo(source_annot_dir: str, source_image_dir: str, target_annot_dir: str, low_dim_cutoff: Optional[int], low_area_cutoff: Optional[float]):
+ """
+ low_dim_cutoff: int, lower cutoff for bounding boxes width/height dims in pixels
+ low_area_cutoff: float, lower area perc cutoff for bounding box areas in perc
+ """
+ if not all([osp.isdir(source_annot_dir), osp.isdir(source_image_dir)]):
+ raise ValueError(f"source_annot_dir and source_image_dir must be directories")
+ src_annot_path = osp.join(source_annot_dir, "*")
+ src_image_path = osp.join(source_image_dir, "*")
+ src_annot_paths = sorted(glob.glob(src_annot_path))
+ src_image_paths = [p for p in sorted(glob.glob(src_image_path)) if osp.splitext(p)[-1] in IMG_EXT]
+ assert len(src_image_paths) == len(src_annot_paths)
+
+ os.makedirs(target_annot_dir, exist_ok=True)
+ low_dim_cutoff = float('-inf') if not low_dim_cutoff else low_dim_cutoff
+ low_area_cutoff = float('-inf') if not low_area_cutoff else low_area_cutoff
+ target_img_list_fpath = osp.join(osp.dirname(target_annot_dir), source_annot_dir.split('/')[-2].lower()+".txt")
+
+ with tqdm.tqdm(total=len(src_image_paths)) as pbar, open(target_img_list_fpath, "w") as imgw:
+ orig_box_count = new_box_count = 0
+ for src_annot_file, src_image_file in zip(src_annot_paths, src_image_paths):
+ try:
+ iw, ih = imagesize.get(src_image_file)
+ target_annot_file = osp.join(target_annot_dir, src_annot_file.split('/')[-1])
+ with open(src_annot_file, "r") as fr, open(target_annot_file, "w") as fw:
+ for coords in fr:
+ annots = list(map(int, coords.strip().strip(',').split(',')))
+ x, y = annots[0], annots[1]
+ w, h = annots[2], annots[3]
+ score, class_id, occu = annots[4], annots[5], annots[7]
+ if class_id not in CLASS_2_CONSIDER: # only keep classes to consider
+ continue
+ orig_box_count += 1
+ if w < low_dim_cutoff or h < low_dim_cutoff: # cutoff value for dims to remove outliers
+ continue
+ area_perc = 100 * (w * h) / (iw * ih)
+ if area_perc < low_area_cutoff: # cutoff value for area perc to remove outliers
+ continue
+
+ xc, yc = x + (w / 2), y + (h / 2)
+ # only use objects used for eval along and all levels of occlusion (0,1,2)
+ if score and occu <= 2:
+ class_id = CLASS_ID_REMAP[class_id] if CLASS_ID_REMAP else class_id
+ fw.write(f"{class_id} {xc/iw} {yc/ih} {w/iw} {h/ih}\n")
+ new_box_count += 1
+ target_image_path = osp.join(osp.dirname(osp.dirname(target_annot_file)), "images",
+ osp.basename(target_annot_file).split('.')[0] + osp.splitext(src_image_file)[1])
+ imgw.write(f"{osp.abspath(target_image_path)}\n")
+ except Exception as excep:
+ print(f"{excep}: Error reading img {src_image_file}")
+ pbar.update(1)
+ print(f"Original Box Count: {orig_box_count}. Converted Box Count {new_box_count}")
+ print(f"{100 * (new_box_count) / orig_box_count:.2f}% of total boxes kept")
+
+
+def main():
+ args = get_parsed_args()
+ print(args)
+ conv_visdrone_2_yolo(args.source_annot_dir, args.source_image_dir, args.target_annot_dir,
+ args.low_dim_cutoff, args.low_area_cutoff)
+
+
+if __name__ == "__main__":
+ main()

adv_patch_gen/conv_visdrone_2_yolo/disp_visdrone.py

@@ -0,0 +1,54 @@
+import argparse
+import sys
+sys.path.append(".")
+
+import cv2
+from adv_patch_gen.conv_visdrone_2_yolo.utils import get_annot_img_paths, load_visdrone_annots_as_np
+
+
+# visdrone dataset classes
+# ignore(0), pedestrian(1), people(2), bicycle(3),
+# car(4), van(5), truck(6), tricycle(7), awning-tricycle(8),
+# bus(9), motor(10), others(11)
+
+ANNOT_EXT = {".txt"}
+IMG_EXT = {".jpg", ".png"}
+CLASS_IDS_2_CONSIDER = {4, 5, 6, 9}
+
+
+def get_parsed_args():
+ parser = argparse.ArgumentParser(
+ description="Disp VisDrone annotated images")
+ parser.add_argument('-a', '--source_annot_dir', type=str, dest="source_annot_dir", required=True,
+ help='VisDrone annotation source dir. Should contain annot txt files (default: %(default)s)')
+ parser.add_argument('-i', '--source_image_dir', type=str, dest="source_image_dir", required=True,
+ help='VisDrone images source dir. Should contain image files (default: %(default)s)')
+ args = parser.parse_args()
+ return args
+
+
+def main():
+ args = get_parsed_args()
+ annot_paths, image_paths = get_annot_img_paths(
+ args.source_annot_dir, args.source_image_dir, ANNOT_EXT, IMG_EXT)
+
+ for ant, img in zip(annot_paths, image_paths):
+ image = cv2.imread(img)
+ annots = load_visdrone_annots_as_np(ant)
+ for annot in annots:
+ score, class_id = annot[4], annot[5]
+ if class_id not in CLASS_IDS_2_CONSIDER: # car, van ,truck, bus
+ continue
+ color = (0, 0, 255) if score == 0 else (0, 255, 0)
+ x1, y1, x2, y2 = annot[:4]
+ cv2.rectangle(image, (x1, y1), (x2, y2), color, 1)
+
+ cv2.imshow("VisDrone annot visualized", image)
+ key = cv2.waitKey()
+ if key == ord("q"):
+ cv2.destroyAllWindows()
+ exit()
+
+
+if __name__ == "__main__":
+ main()

adv_patch_gen/conv_visdrone_2_yolo/disp_yolo.py

@@ -0,0 +1,59 @@
+import argparse
+import sys
+sys.path.append(".")
+
+import cv2
+from adv_patch_gen.conv_visdrone_2_yolo.utils import get_annot_img_paths, load_yolo_annots_as_np
+
+
+# visdrone dataset classes
+# ignore(0), pedestrian(1), people(2), bicycle(3),
+# car(4), van(5), truck(6), tricycle(7), awning-tricycle(8),
+# bus(9), motor(10), others(11)
+# CLASS_IDS_2_CONSIDER = {4, 5, 6, 9}
+
+ANNOT_EXT = {".txt"}
+IMG_EXT = {".jpg", ".png"}
+# custom yolo classes
+# car(0), van(1), truck(2), bus(3)
+CLASS_IDS_2_CONSIDER = {0, 1, 2, 3}
+
+
+def get_parsed_args():
+ parser = argparse.ArgumentParser(
+ description="Disp YOLO annotated images")
+ parser.add_argument('-a', '--source_annot_dir', type=str, dest="source_annot_dir", required=True,
+ help='YOLO annotation source dir. Should contain annot txt files (default: %(default)s)')
+ parser.add_argument('-i', '--source_image_dir', type=str, dest="source_image_dir", required=True,
+ help='YOLO images source dir. Should contain image files (default: %(default)s)')
+ args = parser.parse_args()
+ return args
+
+
+def main():
+ args = get_parsed_args()
+ annot_paths, image_paths = get_annot_img_paths(
+ args.source_annot_dir, args.source_image_dir, ANNOT_EXT, IMG_EXT)
+
+ for ant, img in zip(annot_paths, image_paths):
+ image = cv2.imread(img)
+ annots = load_yolo_annots_as_np(ant)
+ for annot in annots:
+ class_id = annot[0]
+ if class_id not in CLASS_IDS_2_CONSIDER: # car, van ,truck, bus
+ continue
+ ih, iw, _ = image.shape
+ xc, yc, w, h = annot[1] * iw, annot[2] * ih, annot[3] * iw, annot[4] * ih
+ x1, y1, x2, y2 = xc - w / 2, yc - h / 2, xc + w / 2, yc + h / 2
+ x1, y1, x2, y2 = map(int, (x1, y1, x2, y2))
+ cv2.rectangle(image, (x1, y1), (x2, y2), (0, 0, 255), 1)
+
+ cv2.imshow("YOLO annot visualized", image)
+ key = cv2.waitKey()
+ if key == ord("q"):
+ cv2.destroyAllWindows()
+ exit()
+
+
+if __name__ == "__main__":
+ main()

adv_patch_gen/conv_visdrone_2_yolo/utils.py

@@ -0,0 +1,39 @@
+import os
+import glob
+from typing import List, Tuple, Set
+
+import numpy as np
+
+
+def get_annot_img_paths(annot_dir: str, image_dir: str, annot_ext: Set[str], img_ext: Set[str]) -> Tuple[List[str], List[str]]:
+ annots_path = os.path.join(annot_dir, "*")
+ images_path = os.path.join(image_dir, "*")
+ annot_paths = [p for p in sorted(glob.glob(annots_path)) if os.path.splitext(p)[-1] in annot_ext]
+ image_paths = [p for p in sorted(glob.glob(images_path)) if os.path.splitext(p)[-1] in img_ext]
+
+ assert len(annot_paths) == len(image_paths)
+ return annot_paths, image_paths
+
+
+def load_visdrone_annots_as_np(annot_file: str) -> np.ndarray:
+ annot_list = []
+ with open(annot_file, "r") as f:
+ for values in f:
+ annots = list(map(int, values.strip().strip(',').split(',')))
+ x1, y1 = annots[0], annots[1]
+ x2, y2 = x1 + annots[2], y1 + annots[3]
+ score, class_id, occu = annots[4], annots[5], annots[7]
+ annot_list.append([x1, y1, x2, y2, score, class_id, occu])
+ return np.asarray(annot_list)
+
+
+def load_yolo_annots_as_np(annot_file: str) -> np.ndarray:
+ annot_list = []
+ with open(annot_file, "r") as f:
+ for values in f:
+ annots = list(map(float, values.strip().split()))
+ class_id = annots[0]
+ xc, yc = annots[1], annots[2]
+ w, h = annots[3], annots[4]
+ annot_list.append([class_id, xc, yc, w, h])
+ return np.asarray(annot_list)

adv_patch_gen/test_notebooks/30_rgb_triplets.csv

@@ -0,0 +1,30 @@
+0.10588,0.054902,0.1098
+0.48235,0.094118,0.16863
+0.50196,0.52549,0.17647
+0.082353,0.31765,0.18431
+0.47843,0.61176,0.51765
+0.07451,0.3098,0.45882
+0.67843,0.14902,0.18039
+0.086275,0.14118,0.26275
+0.26667,0.36863,0.47843
+0.76078,0.54118,0.5451
+0.73333,0.49412,0.27451
+0.25882,0.35294,0.18039
+0.47843,0.22353,0.36471
+0.27059,0.086275,0.11765
+0.7098,0.32157,0.2
+0.27451,0.13725,0.29412
+0.75294,0.75686,0.63137
+0.28627,0.54902,0.41176
+0.47451,0.2902,0.15294
+0.74902,0.70196,0.28627
+0.098039,0.42745,0.44314
+0.50588,0.65098,0.65882
+0.12549,0.42745,0.23529
+0.4902,0.58431,0.33725
+0.26275,0.49412,0.26275
+0.07451,0.14902,0.12549
+0.090196,0.20392,0.36078
+0.68627,0.15686,0.30196
+0.30196,0.5451,0.57647
+0.71765,0.32941,0.40784