admin/momentry_core
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: schema tracking, SHA256 integrity, identity UUID fix, 3-angle face match, cuts table, trace stranger_id

Browse Source
This commit is contained in:
Accusys
2026-05-16 03:10:50 +08:00
parent c41f7e0c6e
commit 5317cb4bec
13 changed files with 242 additions and 80 deletions
Download Patch File Download Diff File Expand all files Collapse all files

94
src/api/identity_agent_api.rs
View File

@@ -676,12 +676,12 @@ async fn match_faces_iterative(pool: &sqlx::PgPool, file_uuid: &str) -> anyhow::
tmdb_rows.len()
);
// Step 2: 載入所有 face_detections，按 trace_id 分組
// Step 2: 載入所有 face_detections（含 frame_number），按 trace_id 分組
let fd_table = schema::table_name("face_detections");
let fd_rows = sqlx::query_as::<_, (i32, Vec<f32>)>(
&format!("SELECT trace_id, embedding FROM {} \
let fd_rows = sqlx::query_as::<_, (i32, i32, Vec<f32>)>(
&format!("SELECT trace_id, frame_number, embedding FROM {} \
WHERE file_uuid=$1 AND trace_id IS NOT NULL AND embedding IS NOT NULL \
ORDER BY trace_id", fd_table),
ORDER BY trace_id, frame_number", fd_table),
)
.bind(file_uuid)
.fetch_all(pool)
@@ -692,27 +692,38 @@ async fn match_faces_iterative(pool: &sqlx::PgPool, file_uuid: &str) -> anyhow::
return Ok(0);
}
// 分組：trace_id → Vec<embedding>
// 分組：trace_id → (frame_number, embedding)
use std::collections::HashMap;
let mut trace_faces: HashMap<i32, Vec<Vec<f32>>> = HashMap::new();
for (tid, emb) in &fd_rows {
trace_faces
let mut trace_faces_raw: HashMap<i32, Vec<(i32, Vec<f32>)>> = HashMap::new();
for (tid, frame, emb) in &fd_rows {
trace_faces_raw
.entry(*tid)
.or_insert_with(Vec::new)
.push(emb.clone());
.push((*frame, emb.clone()));
}
// 去重：同一個 trace 內，embedding 太接近的只留一個
for faces in trace_faces.values_mut() {
faces.sort_by(|a, b| b[0].partial_cmp(&a[0]).unwrap_or(std::cmp::Ordering::Equal));
faces.dedup_by(|a, b| cosine_similarity(a, b) > 0.99);
// 從每個 trace 選取不同角度的 3 個 face embedding
// 策略：按 frame_number 排序，取前中後各 1 個
let mut trace_samples: HashMap<i32, Vec<Vec<f32>>> = HashMap::new();
for (tid, mut faces) in trace_faces_raw {
faces.sort_by_key(|(frame, _)| *frame);
let n = faces.len();
let indices = if n <= 3 {
(0..n).collect()
} else {
let mid = n / 2;
vec![0, mid, n - 1]
};
let samples: Vec<Vec<f32>> = indices.iter().map(|&i| faces[i].1.clone()).collect();
trace_samples.insert(tid, samples);
}
let total_traces = trace_faces.len();
let total_traces = trace_samples.len();
let sample_count: usize = trace_samples.values().map(|v| v.len()).sum();
tracing::info!(
"[FaceMatch] Loaded {} traces with {} faces",
"[FaceMatch] Loaded {} traces, sampled {} embeddings (3-angle)",
total_traces,
fd_rows.len()
sample_count
);
// Step 3: 建立 TMDb 查找表
@@ -722,12 +733,13 @@ async fn match_faces_iterative(pool: &sqlx::PgPool, file_uuid: &str) -> anyhow::
const TH: f32 = 0.50;
let mut matched: HashMap<i32, String> = HashMap::new(); // trace_id → identity_name
// Round 1: 直接比對 TMDb
for (&tid, faces) in &trace_faces {
// Round 1: 用 3-angle samples 比對 TMDb
// 每個 trace 選 3 個不同角度 face，取最高 similarity
for (&tid, samples) in &trace_samples {
let mut best_name = String::new();
let mut best_sim = 0.0f32;
for (_, ref name, ref tmdb_emb) in &tmdb_seeds {
for face_emb in faces {
for face_emb in samples {
let s = cosine_similarity(face_emb, tmdb_emb);
if s > best_sim {
best_sim = s;
@@ -751,31 +763,33 @@ async fn match_faces_iterative(pool: &sqlx::PgPool, file_uuid: &str) -> anyhow::
// 建立 seed pool: name → Vec<embedding>
let mut seed_pool: HashMap<String, Vec<&Vec<f32>>> = HashMap::new();
for (&tid, name) in &matched {
if let Some(faces) = trace_faces.get(&tid) {
if let Some(samples) = trace_samples.get(&tid) {
seed_pool
.entry(name.clone())
.or_default()
.extend(faces.iter());
.extend(samples.iter());
}
}
let mut new_matches: Vec<(i32, String)> = Vec::new();
for (&tid, faces) in &trace_faces {
for (&tid, samples) in &trace_samples {
if matched.contains_key(&tid) {
continue;
}
let mut best_name = String::new();
let mut best_sim = 0.0f32;
if faces.is_empty() {
if samples.is_empty() {
continue;
}
let ref_face = &faces[0];
// 用 3-angle samples 分別比對 seed，取最高 similarity
for (name, seed_faces) in &seed_pool {
for seed in seed_faces {
let s = cosine_similarity(ref_face, seed);
if s > best_sim {
best_sim = s;
best_name = name.clone();
for face_emb in samples {
for seed in seed_faces {
let s = cosine_similarity(face_emb, seed);
if s > best_sim {
best_sim = s;
best_name = name.clone();
}
}
}
}
@@ -799,7 +813,7 @@ async fn match_faces_iterative(pool: &sqlx::PgPool, file_uuid: &str) -> anyhow::
}
}
// Step 5: 寫入 DB
// Step 5: 寫入 DB — 已匹配的設 identity_id
let identities_table = schema::table_name("identities");
let fd_table = schema::table_name("face_detections");
let mut updated = 0usize;
@@ -823,11 +837,27 @@ async fn match_faces_iterative(pool: &sqlx::PgPool, file_uuid: &str) -> anyhow::
}
}
// Step 6: 未匹配的 trace 設 stranger_id = trace_id
// trace_id 在同一個 file 內是 sequential integer，直接複用為 stranger_id
let stranger_update = sqlx::query(
&format!(
"UPDATE {} SET stranger_id = trace_id \
WHERE file_uuid = $1 AND trace_id IS NOT NULL AND identity_id IS NULL \
AND (stranger_id IS NULL OR stranger_id != trace_id)",
fd_table
)
)
.bind(file_uuid)
.execute(pool)
.await?;
let stranger_count = stranger_update.rows_affected();
tracing::info!(
"[FaceMatch] Done: {}/{} traces matched ({}%)",
"[FaceMatch] Done: {}/{} traces matched ({}%), {} strangers",
matched.len(),
total_traces,
matched.len() * 100 / total_traces
matched.len() * 100 / total_traces,
stranger_count
);
Ok(updated)
}