163 lines
4.9 KiB
Matlab
163 lines
4.9 KiB
Matlab
clear; clc;
|
|
|
|
addpath(genpath('src'));
|
|
addpath(genpath('data'));
|
|
|
|
%% data preparation
|
|
global data_path;
|
|
% data_path = '/your/mulran/sequence/dir/Riverside02/';
|
|
data_path = '/media/user/My Passport/data/MulRan_eval/Riverside_2_20190816/20190816/';
|
|
% ### NOTE: Use this sequence directory structure
|
|
% example:
|
|
% /your/MulRan/sequence/dir/Riverside02/
|
|
% L sensor_data/
|
|
% L radar/
|
|
% L polar/
|
|
% L {unix_times}.png
|
|
% L global_pose.csv
|
|
|
|
down_shape = [40, 120];
|
|
[data_scancontexts, data_ringkeys, data_poses] = loadData(down_shape);
|
|
|
|
|
|
%% main - global recognizer
|
|
revisit_criteria = 5; % in meter (recommend test for 5, 10, 20 meters)
|
|
keyframe_gap = 1; % for_fast_eval (if 1, no skip)
|
|
|
|
global num_candidates; num_candidates = 5;
|
|
% NOTE about num_candidates
|
|
% - 50 was used in the MulRan paper
|
|
% - But we found, interestingly, using less keys showed similar
|
|
% performance - also we can save the computation time of course.
|
|
% - That means our ring key has good disriminative power.
|
|
|
|
global num_node_enough_apart; num_node_enough_apart = 50;
|
|
|
|
% policy (top N)
|
|
num_top_n = 25;
|
|
top_n = linspace(1, num_top_n, num_top_n);
|
|
|
|
% Entropy thresholds
|
|
middle_thres = 0.01;
|
|
thresholds1 = linspace(0, middle_thres, 50);
|
|
thresholds2 = linspace(middle_thres, 1, 50);
|
|
thresholds = [thresholds1, thresholds2];
|
|
num_thresholds = length(thresholds);
|
|
|
|
% Main variables to store the result for drawing PR curve
|
|
num_hits = zeros(num_top_n, num_thresholds);
|
|
num_false_alarms = zeros(num_top_n, num_thresholds);
|
|
num_correct_rejections = zeros(num_top_n, num_thresholds);
|
|
num_misses = zeros(num_top_n, num_thresholds);
|
|
|
|
% main
|
|
loop_log = [];
|
|
|
|
exp_poses = [];
|
|
exp_ringkeys = [];
|
|
exp_scancontexts = {};
|
|
|
|
num_queries = length(data_poses);
|
|
for query_idx = 1:num_queries - 1
|
|
|
|
% save to (online) DB
|
|
query_sc = data_scancontexts{query_idx};
|
|
query_rk = data_ringkeys(query_idx, :);
|
|
query_pose = data_poses(query_idx,:);
|
|
|
|
exp_scancontexts{end+1} = query_sc;
|
|
exp_poses = [exp_poses; query_pose];
|
|
exp_ringkeys = [exp_ringkeys; query_rk];
|
|
|
|
if(rem(query_idx, keyframe_gap) ~= 0)
|
|
continue;
|
|
end
|
|
|
|
if( length(exp_scancontexts) < num_node_enough_apart )
|
|
continue;
|
|
end
|
|
|
|
tree = createns(exp_ringkeys(1:end-(num_node_enough_apart-1), :), 'NSMethod', 'kdtree'); % Create object to use in k-nearest neighbor search
|
|
|
|
% revisitness
|
|
[revisitness, how_far_apart] = isRevisitGlobalLoc(query_pose, exp_poses(1:end-(num_node_enough_apart-1), :), revisit_criteria);
|
|
|
|
% find candidates
|
|
candidates = knnsearch(tree, query_rk, 'K', num_candidates);
|
|
|
|
% find the nearest (top 1) via pairwise comparison
|
|
nearest_idx = 0;
|
|
min_dist = inf; % initialization
|
|
for ith_candidate = 1:length(candidates)
|
|
candidate_node_idx = candidates(ith_candidate);
|
|
candidate_img = exp_scancontexts{candidate_node_idx};
|
|
|
|
% if( abs(query_idx - candidate_node_idx) < num_node_enough_apart)
|
|
% continue;
|
|
% end
|
|
|
|
distance_to_query = sc_dist(query_sc, candidate_img);
|
|
|
|
if( distance_to_query < min_dist)
|
|
nearest_idx = candidate_node_idx;
|
|
min_dist = distance_to_query;
|
|
end
|
|
end
|
|
|
|
% prcurve analysis
|
|
for topk = 1:num_top_n
|
|
for thres_idx = 1:num_thresholds
|
|
threshold = thresholds(thres_idx);
|
|
|
|
reject = 0;
|
|
if( min_dist > threshold)
|
|
reject = 1;
|
|
end
|
|
|
|
if(reject == 1)
|
|
if(revisitness == 0)
|
|
% TN: Correct Rejection
|
|
num_correct_rejections(topk, thres_idx) = num_correct_rejections(topk, thres_idx) + 1;
|
|
else
|
|
% FN: MISS
|
|
num_misses(topk, thres_idx) = num_misses(topk, thres_idx) + 1;
|
|
end
|
|
else
|
|
% if under the theshold, it is considered seen.
|
|
% and then check the correctness
|
|
if( dist_btn_pose(query_pose, exp_poses(nearest_idx, :)) < revisit_criteria)
|
|
% TP: Hit
|
|
num_hits(topk, thres_idx) = num_hits(topk, thres_idx) + 1;
|
|
else
|
|
% FP: False Alarm
|
|
num_false_alarms(topk, thres_idx) = num_false_alarms(topk, thres_idx) + 1;
|
|
end
|
|
end
|
|
|
|
end
|
|
end
|
|
|
|
if( rem(query_idx, 100) == 0)
|
|
disp( strcat(num2str(query_idx/num_queries * 100), ' % processed') );
|
|
end
|
|
|
|
end
|
|
|
|
|
|
%% save the log
|
|
savePath = strcat("pr_result/within ", num2str(revisit_criteria), "m/");
|
|
if((~7==exist(savePath,'dir')))
|
|
mkdir(savePath);
|
|
end
|
|
save(strcat(savePath, 'nCorrectRejections.mat'), 'num_correct_rejections');
|
|
save(strcat(savePath, 'nMisses.mat'), 'num_misses');
|
|
save(strcat(savePath, 'nHits.mat'), 'num_hits');
|
|
save(strcat(savePath, 'nFalseAlarms.mat'), 'num_false_alarms');
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|