Improved group merging.
This commit is contained in:
Daniel
278
media/RESULTS.md
278
media/RESULTS.md
@ -5998,58 +5998,58 @@ Results of the model in various experiments on different datasets.
|
||||
(volleyball)
|
||||
```json
|
||||
{
|
||||
"avg_time_2d": 0.3441050417788394,
|
||||
"avg_time_3d": 0.0033576402578267966,
|
||||
"avg_fps": 2.878006910378003
|
||||
"avg_time_2d": 0.3424515831577885,
|
||||
"avg_time_3d": 0.003363656568097639,
|
||||
"avg_fps": 2.8917175564404274
|
||||
}
|
||||
{
|
||||
"person_nums": {
|
||||
"total_frames": 121,
|
||||
"total_labels": 484,
|
||||
"total_preds": 633,
|
||||
"total_preds": 634,
|
||||
"considered_empty": 0,
|
||||
"valid_preds": 483,
|
||||
"invalid_preds": 150,
|
||||
"invalid_preds": 151,
|
||||
"missing": 1,
|
||||
"invalid_fraction": 0.23697,
|
||||
"precision": 0.76303,
|
||||
"invalid_fraction": 0.23817,
|
||||
"precision": 0.76183,
|
||||
"recall": 0.99793,
|
||||
"f1": 0.86482,
|
||||
"non_empty": 633
|
||||
"f1": 0.86404,
|
||||
"non_empty": 634
|
||||
},
|
||||
"mpjpe": {
|
||||
"count": 483,
|
||||
"mean": 0.035428,
|
||||
"median": 0.032584,
|
||||
"std": 0.013121,
|
||||
"sem": 0.000598,
|
||||
"min": 0.017908,
|
||||
"max": 0.12169,
|
||||
"recall-0.025": 0.216942,
|
||||
"recall-0.05": 0.876033,
|
||||
"recall-0.1": 0.993802,
|
||||
"mean": 0.03745,
|
||||
"median": 0.035063,
|
||||
"std": 0.014194,
|
||||
"sem": 0.000647,
|
||||
"min": 0.01794,
|
||||
"max": 0.136715,
|
||||
"recall-0.025": 0.163223,
|
||||
"recall-0.05": 0.842975,
|
||||
"recall-0.1": 0.991736,
|
||||
"recall-0.15": 0.997934,
|
||||
"recall-0.25": 0.997934,
|
||||
"recall-0.5": 0.997934,
|
||||
"num_labels": 484,
|
||||
"ap-0.025": 0.041041,
|
||||
"ap-0.05": 0.626763,
|
||||
"ap-0.1": 0.797009,
|
||||
"ap-0.15": 0.803699,
|
||||
"ap-0.25": 0.803699,
|
||||
"ap-0.5": 0.803699
|
||||
"ap-0.025": 0.023957,
|
||||
"ap-0.05": 0.580664,
|
||||
"ap-0.1": 0.79534,
|
||||
"ap-0.15": 0.804355,
|
||||
"ap-0.25": 0.804355,
|
||||
"ap-0.5": 0.804355
|
||||
},
|
||||
"head": {
|
||||
"count": 483,
|
||||
"mean": 0.036562,
|
||||
"median": 0.030386,
|
||||
"std": 0.022881,
|
||||
"sem": 0.001042,
|
||||
"mean": 0.037576,
|
||||
"median": 0.030619,
|
||||
"std": 0.023474,
|
||||
"sem": 0.001069,
|
||||
"min": 0.004382,
|
||||
"max": 0.136584,
|
||||
"recall-0.025": 0.392562,
|
||||
"recall-0.05": 0.739669,
|
||||
"recall-0.1": 0.979339,
|
||||
"recall-0.025": 0.378099,
|
||||
"recall-0.05": 0.721074,
|
||||
"recall-0.1": 0.977273,
|
||||
"recall-0.15": 0.997934,
|
||||
"recall-0.25": 0.997934,
|
||||
"recall-0.5": 0.997934,
|
||||
@ -6057,15 +6057,15 @@ Results of the model in various experiments on different datasets.
|
||||
},
|
||||
"shoulder_left": {
|
||||
"count": 483,
|
||||
"mean": 0.034722,
|
||||
"median": 0.029123,
|
||||
"std": 0.020485,
|
||||
"sem": 0.000933,
|
||||
"mean": 0.039129,
|
||||
"median": 0.032213,
|
||||
"std": 0.022975,
|
||||
"sem": 0.001047,
|
||||
"min": 0.004214,
|
||||
"max": 0.142662,
|
||||
"recall-0.025": 0.369835,
|
||||
"recall-0.05": 0.81405,
|
||||
"recall-0.1": 0.987603,
|
||||
"recall-0.025": 0.31405,
|
||||
"recall-0.05": 0.745868,
|
||||
"recall-0.1": 0.979339,
|
||||
"recall-0.15": 0.997934,
|
||||
"recall-0.25": 0.997934,
|
||||
"recall-0.5": 0.997934,
|
||||
@ -6073,111 +6073,111 @@ Results of the model in various experiments on different datasets.
|
||||
},
|
||||
"shoulder_right": {
|
||||
"count": 483,
|
||||
"mean": 0.031759,
|
||||
"median": 0.026109,
|
||||
"std": 0.018407,
|
||||
"sem": 0.000838,
|
||||
"min": 0.00347,
|
||||
"max": 0.110679,
|
||||
"recall-0.025": 0.466942,
|
||||
"recall-0.05": 0.826446,
|
||||
"recall-0.1": 0.993802,
|
||||
"recall-0.15": 0.997934,
|
||||
"mean": 0.035508,
|
||||
"median": 0.030929,
|
||||
"std": 0.021175,
|
||||
"sem": 0.000965,
|
||||
"min": 0.003681,
|
||||
"max": 0.180338,
|
||||
"recall-0.025": 0.384298,
|
||||
"recall-0.05": 0.778926,
|
||||
"recall-0.1": 0.987603,
|
||||
"recall-0.15": 0.995868,
|
||||
"recall-0.25": 0.997934,
|
||||
"recall-0.5": 0.997934,
|
||||
"num_labels": 484
|
||||
},
|
||||
"elbow_left": {
|
||||
"count": 483,
|
||||
"mean": 0.035706,
|
||||
"median": 0.029341,
|
||||
"std": 0.024522,
|
||||
"sem": 0.001117,
|
||||
"min": 0.004211,
|
||||
"mean": 0.039093,
|
||||
"median": 0.031479,
|
||||
"std": 0.027871,
|
||||
"sem": 0.001269,
|
||||
"min": 0.005996,
|
||||
"max": 0.237523,
|
||||
"recall-0.025": 0.386364,
|
||||
"recall-0.05": 0.77686,
|
||||
"recall-0.1": 0.981405,
|
||||
"recall-0.15": 0.993802,
|
||||
"recall-0.025": 0.347107,
|
||||
"recall-0.05": 0.737603,
|
||||
"recall-0.1": 0.966942,
|
||||
"recall-0.15": 0.987603,
|
||||
"recall-0.25": 0.997934,
|
||||
"recall-0.5": 0.997934,
|
||||
"num_labels": 484
|
||||
},
|
||||
"elbow_right": {
|
||||
"count": 483,
|
||||
"mean": 0.03838,
|
||||
"median": 0.031125,
|
||||
"std": 0.034151,
|
||||
"sem": 0.001556,
|
||||
"mean": 0.041455,
|
||||
"median": 0.033784,
|
||||
"std": 0.038189,
|
||||
"sem": 0.001739,
|
||||
"min": 0.003944,
|
||||
"max": 0.362552,
|
||||
"recall-0.025": 0.407025,
|
||||
"recall-0.05": 0.774793,
|
||||
"recall-0.1": 0.964876,
|
||||
"recall-0.15": 0.981405,
|
||||
"max": 0.443462,
|
||||
"recall-0.025": 0.367769,
|
||||
"recall-0.05": 0.741736,
|
||||
"recall-0.1": 0.954545,
|
||||
"recall-0.15": 0.979339,
|
||||
"recall-0.25": 0.993802,
|
||||
"recall-0.5": 0.997934,
|
||||
"num_labels": 484
|
||||
},
|
||||
"wrist_left": {
|
||||
"count": 483,
|
||||
"mean": 0.047139,
|
||||
"median": 0.03772,
|
||||
"std": 0.035565,
|
||||
"sem": 0.00162,
|
||||
"min": 0.003421,
|
||||
"max": 0.314461,
|
||||
"recall-0.025": 0.274793,
|
||||
"recall-0.05": 0.669421,
|
||||
"recall-0.1": 0.923554,
|
||||
"recall-0.15": 0.979339,
|
||||
"mean": 0.04895,
|
||||
"median": 0.039104,
|
||||
"std": 0.036788,
|
||||
"sem": 0.001676,
|
||||
"min": 0.002618,
|
||||
"max": 0.309556,
|
||||
"recall-0.025": 0.258264,
|
||||
"recall-0.05": 0.650826,
|
||||
"recall-0.1": 0.913223,
|
||||
"recall-0.15": 0.977273,
|
||||
"recall-0.25": 0.995868,
|
||||
"recall-0.5": 0.997934,
|
||||
"num_labels": 484
|
||||
},
|
||||
"wrist_right": {
|
||||
"count": 482,
|
||||
"mean": 0.047946,
|
||||
"median": 0.036566,
|
||||
"std": 0.041584,
|
||||
"sem": 0.001896,
|
||||
"min": 0.002446,
|
||||
"max": 0.39908,
|
||||
"recall-0.025": 0.289256,
|
||||
"recall-0.05": 0.652893,
|
||||
"recall-0.1": 0.919421,
|
||||
"recall-0.15": 0.969008,
|
||||
"count": 481,
|
||||
"mean": 0.050009,
|
||||
"median": 0.040081,
|
||||
"std": 0.040668,
|
||||
"sem": 0.001856,
|
||||
"min": 0.003069,
|
||||
"max": 0.35488,
|
||||
"recall-0.025": 0.262397,
|
||||
"recall-0.05": 0.63843,
|
||||
"recall-0.1": 0.900826,
|
||||
"recall-0.15": 0.964876,
|
||||
"recall-0.25": 0.987603,
|
||||
"recall-0.5": 0.995868,
|
||||
"recall-0.5": 0.993802,
|
||||
"num_labels": 484
|
||||
},
|
||||
"hip_left": {
|
||||
"count": 483,
|
||||
"mean": 0.038758,
|
||||
"median": 0.036117,
|
||||
"std": 0.019053,
|
||||
"sem": 0.000868,
|
||||
"min": 0.005977,
|
||||
"max": 0.133493,
|
||||
"recall-0.025": 0.239669,
|
||||
"recall-0.05": 0.774793,
|
||||
"recall-0.1": 0.985537,
|
||||
"recall-0.15": 0.997934,
|
||||
"mean": 0.040026,
|
||||
"median": 0.036626,
|
||||
"std": 0.021724,
|
||||
"sem": 0.00099,
|
||||
"min": 0.004385,
|
||||
"max": 0.191835,
|
||||
"recall-0.025": 0.231405,
|
||||
"recall-0.05": 0.756198,
|
||||
"recall-0.1": 0.979339,
|
||||
"recall-0.15": 0.995868,
|
||||
"recall-0.25": 0.997934,
|
||||
"recall-0.5": 0.997934,
|
||||
"num_labels": 484
|
||||
},
|
||||
"hip_right": {
|
||||
"count": 483,
|
||||
"mean": 0.040711,
|
||||
"median": 0.033421,
|
||||
"std": 0.026405,
|
||||
"sem": 0.001203,
|
||||
"min": 0.005419,
|
||||
"mean": 0.042645,
|
||||
"median": 0.034828,
|
||||
"std": 0.027596,
|
||||
"sem": 0.001257,
|
||||
"min": 0.004215,
|
||||
"max": 0.149838,
|
||||
"recall-0.025": 0.291322,
|
||||
"recall-0.05": 0.75,
|
||||
"recall-0.1": 0.958678,
|
||||
"recall-0.025": 0.27686,
|
||||
"recall-0.05": 0.71281,
|
||||
"recall-0.1": 0.948347,
|
||||
"recall-0.15": 0.997934,
|
||||
"recall-0.25": 0.997934,
|
||||
"recall-0.5": 0.997934,
|
||||
@ -6185,14 +6185,14 @@ Results of the model in various experiments on different datasets.
|
||||
},
|
||||
"knee_left": {
|
||||
"count": 483,
|
||||
"mean": 0.025004,
|
||||
"median": 0.0211,
|
||||
"std": 0.018866,
|
||||
"sem": 0.000859,
|
||||
"min": 0.001613,
|
||||
"mean": 0.025053,
|
||||
"median": 0.021076,
|
||||
"std": 0.018987,
|
||||
"sem": 0.000865,
|
||||
"min": 0.001403,
|
||||
"max": 0.27912,
|
||||
"recall-0.025": 0.595041,
|
||||
"recall-0.05": 0.93595,
|
||||
"recall-0.025": 0.603306,
|
||||
"recall-0.05": 0.923554,
|
||||
"recall-0.1": 0.993802,
|
||||
"recall-0.15": 0.995868,
|
||||
"recall-0.25": 0.995868,
|
||||
@ -6201,14 +6201,14 @@ Results of the model in various experiments on different datasets.
|
||||
},
|
||||
"knee_right": {
|
||||
"count": 483,
|
||||
"mean": 0.025695,
|
||||
"median": 0.022859,
|
||||
"std": 0.014864,
|
||||
"sem": 0.000677,
|
||||
"mean": 0.026099,
|
||||
"median": 0.022868,
|
||||
"std": 0.015211,
|
||||
"sem": 0.000693,
|
||||
"min": 0.001415,
|
||||
"max": 0.094234,
|
||||
"recall-0.025": 0.576446,
|
||||
"recall-0.05": 0.92562,
|
||||
"recall-0.025": 0.557851,
|
||||
"recall-0.05": 0.919421,
|
||||
"recall-0.1": 0.997934,
|
||||
"recall-0.15": 0.997934,
|
||||
"recall-0.25": 0.997934,
|
||||
@ -6217,15 +6217,15 @@ Results of the model in various experiments on different datasets.
|
||||
},
|
||||
"ankle_left": {
|
||||
"count": 483,
|
||||
"mean": 0.028829,
|
||||
"median": 0.023676,
|
||||
"std": 0.033788,
|
||||
"sem": 0.001539,
|
||||
"min": 0.001752,
|
||||
"mean": 0.029784,
|
||||
"median": 0.023999,
|
||||
"std": 0.034378,
|
||||
"sem": 0.001566,
|
||||
"min": 0.002215,
|
||||
"max": 0.497796,
|
||||
"recall-0.025": 0.530992,
|
||||
"recall-0.05": 0.92562,
|
||||
"recall-0.1": 0.981405,
|
||||
"recall-0.05": 0.909091,
|
||||
"recall-0.1": 0.979339,
|
||||
"recall-0.15": 0.991736,
|
||||
"recall-0.25": 0.993802,
|
||||
"recall-0.5": 0.997934,
|
||||
@ -6233,14 +6233,14 @@ Results of the model in various experiments on different datasets.
|
||||
},
|
||||
"ankle_right": {
|
||||
"count": 483,
|
||||
"mean": 0.028136,
|
||||
"median": 0.02508,
|
||||
"std": 0.016445,
|
||||
"sem": 0.000749,
|
||||
"mean": 0.029202,
|
||||
"median": 0.026139,
|
||||
"std": 0.017443,
|
||||
"sem": 0.000795,
|
||||
"min": 0.001964,
|
||||
"max": 0.103825,
|
||||
"recall-0.025": 0.495868,
|
||||
"recall-0.05": 0.890496,
|
||||
"recall-0.025": 0.464876,
|
||||
"recall-0.05": 0.880165,
|
||||
"recall-0.1": 0.993802,
|
||||
"recall-0.15": 0.997934,
|
||||
"recall-0.25": 0.997934,
|
||||
@ -6249,18 +6249,18 @@ Results of the model in various experiments on different datasets.
|
||||
},
|
||||
"joint_recalls": {
|
||||
"num_labels": 6292,
|
||||
"recall-0.025": 0.40782,
|
||||
"recall-0.05": 0.8034,
|
||||
"recall-0.1": 0.9733,
|
||||
"recall-0.15": 0.99189,
|
||||
"recall-0.025": 0.38207,
|
||||
"recall-0.05": 0.77718,
|
||||
"recall-0.1": 0.96615,
|
||||
"recall-0.15": 0.99031,
|
||||
"recall-0.25": 0.99603,
|
||||
"recall-0.5": 0.99777
|
||||
"recall-0.5": 0.99762
|
||||
}
|
||||
}
|
||||
{
|
||||
"total_parts": 6776,
|
||||
"correct_parts": 6740,
|
||||
"pcp": 0.994687
|
||||
"correct_parts": 6729,
|
||||
"pcp": 0.993064
|
||||
}
|
||||
```
|
||||
|
||||
|
||||
@ -1220,8 +1220,7 @@ std::vector<std::tuple<cv::Point3f, cv::Mat, std::vector<int>>> TriangulatorInte
|
||||
cv::Mat &group_pose = std::get<1>(group);
|
||||
|
||||
// Calculate average joint distance
|
||||
float dist_sum = 0.0;
|
||||
size_t count = 0;
|
||||
std::vector<float> dists;
|
||||
for (size_t row = 0; row < num_joints; ++row)
|
||||
{
|
||||
const float *pose_3d_ptr = pose_3d.ptr<float>(row);
|
||||
@ -1236,15 +1235,20 @@ std::vector<std::tuple<cv::Point3f, cv::Mat, std::vector<int>>> TriangulatorInte
|
||||
float dy = pose_3d_ptr[1] - group_pose_ptr[1];
|
||||
float dz = pose_3d_ptr[2] - group_pose_ptr[2];
|
||||
float dist_sq = dx * dx + dy * dy + dz * dz;
|
||||
dist_sum += std::sqrt(dist_sq);
|
||||
count++;
|
||||
dists.push_back(std::sqrt(dist_sq));
|
||||
}
|
||||
}
|
||||
|
||||
if (count > 0)
|
||||
if (dists.size() >= 5)
|
||||
{
|
||||
// Drop highest value to reduce influence of outliers
|
||||
auto max_it = std::max_element(dists.begin(), dists.end());
|
||||
dists.erase(max_it);
|
||||
}
|
||||
if (dists.size() > 0)
|
||||
{
|
||||
// Check if the average joint distance is close enough
|
||||
float avg_dist = dist_sum / count;
|
||||
float avg_dist = std::accumulate(dists.begin(), dists.end(), 0.0);
|
||||
avg_dist /= static_cast<float>(dists.size());
|
||||
if (avg_dist < max_joint_avg_dist && avg_dist < best_dist)
|
||||
{
|
||||
best_dist = avg_dist;
|
||||
@ -1310,7 +1314,113 @@ std::vector<std::tuple<cv::Point3f, cv::Mat, std::vector<int>>> TriangulatorInte
|
||||
}
|
||||
}
|
||||
|
||||
return groups;
|
||||
// Merge close groups
|
||||
// Depending on the inital group creation, one or more groups can be created that in the end
|
||||
// share the same persons, even if they had a larger distance at the beginning
|
||||
// So merge them similar to the group assignment before
|
||||
std::vector<std::tuple<cv::Point3f, cv::Mat, std::vector<int>>> merged_groups;
|
||||
for (size_t i = 0; i < groups.size(); ++i)
|
||||
{
|
||||
size_t num_joints = std::get<1>(groups[i]).rows;
|
||||
auto &group = groups[i];
|
||||
auto &group_visible_counts = per_group_visible_counts[i];
|
||||
|
||||
float best_dist = std::numeric_limits<float>::infinity();
|
||||
int best_group = -1;
|
||||
|
||||
for (size_t j = 0; j < merged_groups.size(); ++j)
|
||||
{
|
||||
auto &merged_group = merged_groups[j];
|
||||
|
||||
// Calculate average joint distance
|
||||
std::vector<float> dists;
|
||||
for (size_t row = 0; row < num_joints; ++row)
|
||||
{
|
||||
const float *group_pose_ptr = std::get<1>(group).ptr<float>(row);
|
||||
const float *merged_pose_ptr = std::get<1>(merged_group).ptr<float>(row);
|
||||
|
||||
float score1 = group_pose_ptr[3];
|
||||
float score2 = merged_pose_ptr[3];
|
||||
|
||||
if (score1 > min_score && score2 > min_score)
|
||||
{
|
||||
float dx = group_pose_ptr[0] - merged_pose_ptr[0];
|
||||
float dy = group_pose_ptr[1] - merged_pose_ptr[1];
|
||||
float dz = group_pose_ptr[2] - merged_pose_ptr[2];
|
||||
float dist_sq = dx * dx + dy * dy + dz * dz;
|
||||
dists.push_back(std::sqrt(dist_sq));
|
||||
}
|
||||
}
|
||||
if (dists.size() >= 5)
|
||||
{
|
||||
// Drop highest value to reduce influence of outliers
|
||||
auto max_it = std::max_element(dists.begin(), dists.end());
|
||||
dists.erase(max_it);
|
||||
}
|
||||
if (dists.size() > 0)
|
||||
{
|
||||
// Check if the average joint distance is close enough
|
||||
float avg_dist = std::accumulate(dists.begin(), dists.end(), 0.0);
|
||||
avg_dist /= static_cast<float>(dists.size());
|
||||
if (avg_dist < max_joint_avg_dist && avg_dist < best_dist)
|
||||
{
|
||||
best_dist = avg_dist;
|
||||
best_group = static_cast<int>(j);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (best_group == -1)
|
||||
{
|
||||
// Create a new group
|
||||
merged_groups.push_back(group);
|
||||
}
|
||||
else
|
||||
{
|
||||
// Update existing group
|
||||
auto &merged_group = merged_groups[best_group];
|
||||
cv::Point3f &merged_center = std::get<0>(merged_group);
|
||||
cv::Mat &merged_group_pose = std::get<1>(merged_group);
|
||||
std::vector<int> &merged_group_indices = std::get<2>(merged_group);
|
||||
|
||||
float n_elems1 = static_cast<float>(merged_group_indices.size());
|
||||
float n_elems2 = static_cast<float>(std::get<2>(group).size());
|
||||
float inv1 = n_elems1 / (n_elems1 + n_elems2);
|
||||
float inv2 = n_elems2 / (n_elems1 + n_elems2);
|
||||
|
||||
// Update group center
|
||||
merged_center.x = (merged_center.x * inv1 + std::get<0>(group).x * inv2);
|
||||
merged_center.y = (merged_center.y * inv1 + std::get<0>(group).y * inv2);
|
||||
merged_center.z = (merged_center.z * inv1 + std::get<0>(group).z * inv2);
|
||||
|
||||
// Update group pose
|
||||
for (size_t row = 0; row < num_joints; ++row)
|
||||
{
|
||||
const float *group_pose_ptr = std::get<1>(group).ptr<float>(row);
|
||||
float *merged_pose_ptr = merged_group_pose.ptr<float>(row);
|
||||
|
||||
if (group_pose_ptr[3] > min_score)
|
||||
{
|
||||
float j_elems1 = static_cast<float>(group_visible_counts[row]);
|
||||
float j_elems2 = static_cast<float>(per_group_visible_counts[best_group][row]);
|
||||
float inv1 = j_elems1 / (j_elems1 + j_elems2);
|
||||
float inv2 = j_elems2 / (j_elems1 + j_elems2);
|
||||
|
||||
merged_pose_ptr[0] = (merged_pose_ptr[0] * inv1 + group_pose_ptr[0] * inv2);
|
||||
merged_pose_ptr[1] = (merged_pose_ptr[1] * inv1 + group_pose_ptr[1] * inv2);
|
||||
merged_pose_ptr[2] = (merged_pose_ptr[2] * inv1 + group_pose_ptr[2] * inv2);
|
||||
merged_pose_ptr[3] = (merged_pose_ptr[3] * inv1 + group_pose_ptr[3] * inv2);
|
||||
group_visible_counts[row]++;
|
||||
}
|
||||
}
|
||||
|
||||
// Merge indices
|
||||
merged_group_indices.insert(
|
||||
merged_group_indices.end(), std::get<2>(group).begin(), std::get<2>(group).end());
|
||||
}
|
||||
}
|
||||
|
||||
return merged_groups;
|
||||
}
|
||||
|
||||
// =================================================================================================
|
||||
|
||||
Reference in New Issue
Block a user