Spaces:
Running
Running
File size: 8,175 Bytes
e4b6a37 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 | import os
# import matplotlib.pyplot as plt
# import matplotlib.image as mpimg
import sys
sys.path.append(os.path.abspath(".")) # one level up
import numpy as np
import cv2
import open3d as o3d
from scipy.spatial.transform import Rotation
from utils.lidar import PointCloud
from utils.camera import ImageData
import utils.utils as utils
from FoL.reranking import run_rerank
from natsort import natsorted, index_natsorted
import torch
from tqdm import tqdm
################## set device based on cuda availability #################
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print('CUDA availability: ' + str(torch.cuda.is_available()))
####################### Functions for matching using numpy on CPU or Pytorch on GPU ###################
def getMatchIndsCPU(ft_ref,ft_qry,topK=20,metric='cosine'):
"""
metric: 'euclidean' or 'cosine'
"""
# dMat = cdist(ft_ref,ft_qry,metric)
ft_qry_norm = ft_qry / np.linalg.norm(ft_qry, axis=1, keepdims=True) # Shape (M, N)
ft_ref_norm = ft_ref / np.linalg.norm(ft_ref, axis=1, keepdims=True) # Shape (C, N)
# Step 2: Compute cosine similarity
dMat = 1 - (ft_ref_norm @ ft_qry_norm.T)
mInds = np.argsort(dMat,axis=0)[:topK].squeeze() # shape: K x ft_qry.shape[0]
return mInds, dMat
def getMatchIndsGPU(ft_ref, ft_qry,topK=20, metric='cosine'):
# metric: 'euclidean' or 'cosine'
ft_qry_tensor = torch.Tensor(ft_qry).to(device)
ft_ref_tensor = torch.Tensor(ft_ref).to(device)
if metric == 'euclidean':
# Use torch's cdist for Euclidean distance
dMat = torch.cdist(ft_ref, ft_qry)
elif metric == 'cosine':
# # Normalize both the query and reference tensors
ft_qry_norm = ft_qry_tensor / ft_qry_tensor.norm(dim=1, keepdim=True)
ft_ref_norm = ft_ref_tensor / ft_ref_tensor.norm(dim=1, keepdim=True)
# Compute cosine similarity (1 - cosine similarity for distance)
dMat = 1 - ft_ref_norm @ ft_qry_norm.t()
# Get the indices of the top 5 closest matches
mInds = torch.argsort(dMat, dim=0)[:topK].squeeze()
return mInds, dMat
qry_sets = [
'20210909_124816_v2',
]
ref_sets = [
'20230509_115540_v2',
]
vpr_descs = [
'FoL',
]
img_calib_file = f"./camera_calib.txt"
dist_tolerance = 10 # metres
# qry_idx = 4
# User parameters
location = 'dalby-to-brigalow'
################ Reference filenames and directories #################################
ref_condition = ''
ref_camera_pos = 'front'
ref_timestamps = []
ref_utms = []
ref_img_filenames = []
ref_utm_filenames = []
for ref_set in ref_sets:
print(f"Loading {ref_set}")
ref_root_directory = f"../../Datasets/dalby/{location}"
ref_vpr_root = f"../../Datasets/dalby/{location}/vpr_ftrs/"
ref_image_dir = f"{ref_root_directory}/{ref_set}/{ref_camera_pos}-imgs/"
ref_utm_dir = f"{ref_root_directory}/{ref_set}/utm/"
this_ref_timestamp = [filename.split('.png')[0] for filename in natsorted(os.listdir(ref_image_dir)) if os.path.isfile(ref_image_dir+filename)]
ref_utms = ref_utms+[np.loadtxt(ref_utm_dir+filename) for filename in natsorted(os.listdir(ref_utm_dir)) if os.path.isfile(ref_utm_dir+filename)][55::]
ref_img_filenames = [filename for filename in natsorted(os.listdir(ref_image_dir)) if os.path.isfile(ref_image_dir+filename)]
ref_utm_filenames = np.array([filename for filename in natsorted(os.listdir(ref_utm_dir)) if os.path.isfile(ref_utm_dir+filename)])[:len(os.listdir(ref_utm_dir))-55]
ref_timestamps = ref_timestamps+this_ref_timestamp
ref_utms = np.array(ref_utms)
for vpr_desc in vpr_descs:
all_results = []
first = True
print(f"Loading references")
for ref_set in ref_sets:
print(f"Loading {ref_set} {vpr_desc} descriptors")
ref_root_directory = f"../../Datasets/dalby/{location}"
ref_vpr_root = f"../../Datasets/dalby/{location}/vpr_ftrs/"
ref_image_dir = f"{ref_root_directory}/{ref_set}/{ref_camera_pos}-imgs/"
ref_name_sort_idx = index_natsorted(os.listdir(ref_image_dir))
ref_ftr = np.load(f"{ref_vpr_root}/{ref_set}/{vpr_desc}/qry_feats.npy")
ref_local_ftr = np.load(f"{ref_vpr_root}/{ref_set}/{vpr_desc}/qry_local_feats.npy")
if first:
ref_ftrs = ref_ftr[ref_name_sort_idx]
ref_local_ftrs = ref_local_ftr[ref_name_sort_idx]
first = False
else:
ref_ftrs = np.vstack((ref_ftrs, ref_ftr[ref_name_sort_idx]))
ref_local_ftrs = np.vstack((ref_local_ftrs, ref_local_ftr[ref_name_sort_idx]))
for qry_set in qry_sets:
################ Query filenames and directories #################################
qry_condition = ''
qry_camera_pos = 'front'
qry_root_directory = f"../../Datasets/dalby/{location}"
qry_vpr_root = f"../../Datasets/dalby/{location}/vpr_ftrs/"
qry_image_dir = f"{qry_root_directory}/{qry_set}/{qry_camera_pos}-imgs/"
qry_utm_dir = f"{qry_root_directory}/{qry_set}/utm/"
qry_timestamps = [filename.split('.png')[0] for filename in natsorted(os.listdir(qry_image_dir)) if os.path.isfile(qry_image_dir+filename)]
qry_utms = np.array([np.loadtxt(qry_utm_dir+filename) for filename in natsorted(os.listdir(qry_utm_dir)) if os.path.isfile(qry_utm_dir+filename)])
qry_name_sort_idx = index_natsorted(os.listdir(qry_image_dir))
qry_ftrs = np.load(f"{qry_vpr_root}/{qry_set}/{vpr_desc}/qry_feats.npy")
qry_local_ftrs = np.load(f"{qry_vpr_root}/{qry_set}/{vpr_desc}/qry_local_feats.npy")
qry_ftrs = qry_ftrs[qry_name_sort_idx]
qry_local_ftrs = qry_local_ftrs[qry_name_sort_idx]
# mInds, dMat = getMatchIndsGPU(ref_ftrs,qry_ftrs,topK=1)
# mInds = mInds.cpu().numpy()
mInds = run_rerank(qry_ftrs, ref_ftrs, qry_local_ftrs, ref_local_ftrs, recall_values=[1, 5, 10, 20])[:,0]
in_tol = []
dists = []
valid_qry = 0
qry_utm_timestamps, qry_utm_idxs = utils.get_all_corr_files(qry_timestamps, [qry_utm_dir,])
ref_utm_timestamp, ref_utm_idxs = utils.get_all_corr_files(ref_timestamps, [ref_utm_dir,])
for qry_idx in tqdm(range(len(qry_timestamps))):
qry_image_timestamp = qry_timestamps[qry_idx]
qry_image_filename = f"{qry_image_dir}/{qry_image_timestamp}.png"
qry_utm = qry_utms[qry_utm_idxs[qry_idx]]
diffs = ref_utms - qry_utm # shape (N, 2)
qry_dists = np.linalg.norm(diffs, axis=1) # shape (N,)
if qry_dists.min() > dist_tolerance:
continue
else:
valid_qry += 1
ref_utm = ref_utms[ref_utm_idxs[int(mInds[qry_idx])]]
diff = ref_utm - qry_utm # shape (N, 2)
dist = np.linalg.norm(diff) # shape (N,)
dists.append(dist)
if dist < dist_tolerance:
in_tol.append(1)
else:
in_tol.append(0)
# qry_image = ImageData(qry_image_filename, img_calib_file)
# fig, ax = plt.subplots(1, 2, figsize=(19.4, 6))
# ax[0].clear()
# ax[1].clear()
# ax[0].imshow(qry_image.image[:, :, ::-1])
# ax[0].set_title(f"{qry_image_timestamp}.png")
# ax[0].axis("off")
# # Show matching reference image
# # ref_img_timestamp = utils.get_corr_files(ref_timestamps[int(mInds[qry_idx])], [ref_image_dir,])
# ref_image = ImageData(f"{ref_image_dir}/{ref_timestamps[int(mInds[qry_idx])]}.png", img_calib_file)
# ax[1].imshow(ref_image.image[:, :, ::-1])
# ax[1].set_title(f"{ref_timestamps[int(mInds[qry_idx])]}\nDist={dist:.2f}m")
# ax[1].axis("off")
# fig.canvas.draw()
print(f"Recall for {qry_set} using {vpr_desc}: {np.sum(np.array(in_tol))/valid_qry:.02%}")
all_results.append(np.sum(np.array(in_tol))/valid_qry)
# plt.figure()
# plt.plot(np.clip(dists, 0, 30))
# plt.ylim((0,35))
print(f"All {vpr_desc} results:")
print(all_results) |