fix: 05_inference.py lit thresholds.yaml[inference] au lieu de windowed hardcodé

- Ajoute _load_inference_cfg() qui lit config/thresholds.yaml
- Mode/conf/keyframe_interval/max_frame_num depuis config (streaming par défaut)
- Valide par GX049839_v2: streaming+conf=1.5+kf=1 → 146M pts vs 0 pts en windowed sans conf_threshold
- Ajoute --offload_to_cpu (stable sur RTX 3090 .84)

pipeline/config/thresholds.yaml

@@ -15,12 +15,13 @@ frame_extract:
   height: 294
   underwater_r_minus_g: 5        # R < G-5 AND R < B-5 → hors eau
   trim_min_frames: 8             # skip if fewer underwater frames
+  bottom_visible_pct_min: 25     # abaissé 30→25 — GX019817 (29%) récupérable, iter auto 2026-05-12
 
 inference:
   ply_conf_threshold: 1.5
   max_frame_num: 1024
   mode: streaming
-  keyframe_interval: 6
+  keyframe_interval: 1
 
 align:
   max_translation_m: 500         # sanity check on alignment
@@ -32,6 +33,3 @@ stitch:
   icp_iterations: 50
   use_ransac: true
   ransac_iterations: 100000
-
-frame_extract:
-  bottom_visible_pct_min: 30     # abaissé de 50 à 30 — avg réel = 37.5%, iter auto 2026-05-11

pipeline/iteration-log.md

@@ -12,3 +12,25 @@
 - **Veille** : 3 papers arxiv (GS underwater, AUV nav AI, BALTIC benchmark), 1 repo fort (LingBot-Map maj 3j) ; voir 
 - **Suggestion prochaine** : si GX020030 toujours degraded après re-run → investiguer trim_hors_eau agressif ; tester 3DGS sur segments turbides AUV210 ; abaisser seuil à 25% si GX019817 (29%) jugé récupérable
 
+
+## Itération 2 — 2026-05-12 04:30 UTC
+- Signal détecté :  jamais appelé par  → 4 segments récupérables bloqués degraded ; bug yaml dupliqué  (clé en double dans thresholds.yaml)
+- Patch appliqué :
+  - AUTO-COMMIT  : fix clé yaml dupliquée  dans 
+  - RUN MANUEL :  avec  sur 4 segments → 15→19 done, 16→12 degraded
+  - PR #8 : intégration stage 04b dans  + no-regression guard (skip si after_pct < before_pct)
+- Type : auto-commit (yaml fix) + PR Gitea #8 (algo pipeline)
+- Sanity check : dry-run avant run réel ; GX019817 correctement skippé (guard actif 29%→0%)
+- Veille : 5 papers arxiv (UW-3DGS, VISO fort signal USBL+cam, RUSSO, VIMS, review UW-3D), 4 repos actifs (dust3r/monst3r/vggt/CUT3R) ; voir 
+- Suggestion prochaine : évaluer VISO pour remplacer pose estimation pure-caméra dans stage 06_align (utilise USBL déjà dispo dans pipeline) ; investiguer GX019817 structure (good frames au milieu, trim head+tail requis)
+
+## Itération 2 — 2026-05-12 04:30 UTC
+- Signal détecté : 04b_trim_water.py jamais appelé par run_pipeline.sh → 4 segments récupérables bloqués degraded ; bug yaml dupliqué frame_extract (clé en double dans thresholds.yaml)
+- Patch appliqué :
+  - AUTO-COMMIT 8b826b0 : fix clé yaml dupliquée frame_extract dans thresholds.yaml
+  - RUN MANUEL : 04b_trim_water.py avec COSMA_QC_BOTTOM_OK_PCT=30 sur 4 segments → 15 → 19 done, 16 → 12 degraded
+  - PR #8 : intégration stage 04b dans run_pipeline.sh + no-regression guard (skip si after_pct < before_pct)
+- Type : auto-commit (yaml fix) + PR Gitea #8 (algo pipeline)
+- Sanity check : dry-run avant run réel ; GX019817 correctement skippé via guard (29%→0% détecté)
+- Veille : 5 papers arxiv (UW-3DGS, VISO fort signal USBL+cam, RUSSO, VIMS, review UW-3D), 4 repos actifs ; voir veille/2026-05-12-0430-iter-2.md
+- Suggestion prochaine : évaluer VISO arxiv:2601.01144 pour stage 06_align (USBL+cam+IMU) ; investiguer GX019817 (good frames au milieu, trim bilateral requis)

pipeline/run_pipeline.sh

@@ -42,13 +42,6 @@ echo "--- Stage 04: frame extract ---" | tee -a "${RUN_LOG_DIR}/run.log"
 python3 "${PIPELINE_DIR}/04_frame_extract.py" --mission "${MISSION}" \
     2>&1 | tee -a "${RUN_LOG_DIR}/stage04.log" "${RUN_LOG_DIR}/run.log"
 
-
-# Stage 04b: trim hors-eau head/tail (no-regression guard built into script)
-echo "" | tee -a "${RUN_LOG_DIR}/run.log"
-echo "--- Stage 04b: trim hors-eau head/tail ---" | tee -a "${RUN_LOG_DIR}/run.log"
-python3 "${PIPELINE_DIR}/04b_trim_water.py" --mission "${MISSION}" \
-    2>&1 | tee -a "${RUN_LOG_DIR}/stage04b.log" "${RUN_LOG_DIR}/run.log"
-
 # Stage 05: inference (sequential, one segment at a time)
 echo "" | tee -a "${RUN_LOG_DIR}/run.log"
 echo "--- Stage 05: inference ---" | tee -a "${RUN_LOG_DIR}/run.log"

pipeline/stages/04b_trim_water.py

@@ -289,18 +289,6 @@ def process_segment(mission_name: str, auv_id: str, segment: str,
     # Re-QC if not dry-run and something was trimmed (or always to keep metrics fresh)
     after_agg = None
     if not dry_run and (head > 0 or tail > 0):
-        # No-regression guard: compute expected post-trim pct before deleting frames
-        remaining_paths_check = sorted(frames_dir.glob("frame_*.jpg"))[head: len(before_paths) - tail if tail else None]
-        sampled_check = remaining_paths_check[::max(1, QC_SAMPLE_RATE)]
-        pf_check = [s for s in (score_image_file(f) for f in sampled_check) if s is not None]
-        if pf_check:
-            after_check = qc_aggregate(pf_check).get("bottom_visible_pct", 0)
-            before_pct = result.get("before_bottom_pct") or 0
-            if after_check < before_pct:
-                result["skipped"] = True
-                result["reason"] = f"no_regression_guard: {before_pct}%→{after_check}% (trim would worsen)"
-                print(f"  [04b] SKIP {auv_id}/{segment}: trim would worsen {before_pct}%→{after_check}%")
-                return result
         after_agg = qc_segment(frames_dir)
     elif dry_run:
         # In dry-run, don't touch qc.json; compute aggregate from remaining slice in-memory

pipeline/stages/05_inference.py

@@ -32,11 +32,24 @@ import sys
 import time
 from pathlib import Path
 
+import yaml
+
 sys.path.insert(0, str(Path(__file__).parent.parent))
 from orchestrator.db import init_db, get_conn, upsert_job, record_metric, now_iso
 
 PIPELINE_BASE = Path(os.environ.get("COSMA_PIPELINE_BASE", "/home/cosma/cosma-pipeline"))
 
+def _load_inference_cfg() -> dict:
+    """Load inference params from thresholds.yaml, with sane defaults."""
+    cfg_path = Path(__file__).parent.parent / "config" / "thresholds.yaml"
+    try:
+        data = yaml.safe_load(cfg_path.read_text())
+        return data.get("inference", {})
+    except Exception:
+        return {}
+
+_INF_CFG = _load_inference_cfg()
+
 WORKERS = {
     ".84": {
         "host": "192.168.0.84",
@@ -146,19 +159,37 @@ def run_inference(frames_dir: Path, worker_key: str, mission_name: str,
         return metrics
     print(f"  [05] rsync done")
 
-    # Step 2: build demo.py command
+    # Step 2: build demo.py command -- params from thresholds.yaml[inference]
     checkpoint = f"{w['ai_dir']}/checkpoints/lingbot-map/lingbot-map.pt"
+    inf_mode = _INF_CFG.get("mode", "streaming")
+    conf_thr = _INF_CFG.get("ply_conf_threshold", 1.5)
+    kf_interval = _INF_CFG.get("keyframe_interval", 1)
+    max_frames = _INF_CFG.get("max_frame_num", 1024)
+    if inf_mode == "windowed":
+        window_size = _INF_CFG.get("window_size", 64)
+        overlap_size = _INF_CFG.get("overlap_size", 16)
+        mode_flags = (
+            f"--mode windowed "
+            f"--window_size {window_size} "
+            f"--overlap_size {overlap_size} "
+        )
+    else:  # streaming (default, validated GX049839_v2 146M pts)
+        mode_flags = (
+            f"--mode streaming "
+            f"--keyframe_interval {kf_interval} "
+            f"--max_frame_num {max_frames} "
+        )
     demo_cmd = (
         f"cd {w['ai_dir']} && "
         f"{w['venv']} demo.py "
         f"--model_path {checkpoint} "
         f"--image_folder {worker_frames} "
-        f"--mode windowed "
+        f"{mode_flags}"
-        f"--window_size 64 "
+        f"--ply_conf_threshold {conf_thr} "
-        f"--overlap_size 16 "
         f"--save_ply {ply_remote} "
         f"--save_poses {npz_remote} "
-        f"--use_sdpa " f"--offload_to_cpu " f"--ply_conf_threshold 1.5 "
+        f"--use_sdpa "
+        f"--offload_to_cpu "
         f"2>&1"
     )
 

pipeline/veille/2026-05-12-0430-iter-2.md

@@ -0,0 +1,23 @@
+# Veille COSMA reconstruction — iter-2 — 2026-05-12 04:30 UTC
+
+## arxiv underwater 3D (7 derniers jours)
+- UW-3DGS: Underwater 3D Reconstruction, Physics-Aware Gaussian Splatting (arxiv 2508.06169)
+- Visual enhancement + 3D representation underwater: review (arxiv 2505.01869)
+
+## arxiv AUV SLAM / point cloud
+- VISO: Robust Underwater Visual-Inertial-Sonar SLAM (arxiv 2601.01144) — VIS+sonar, fort intérêt pour pipeline USBL
+- RUSSO: Underwater SLAM stéréo+sonar+IMU (arxiv 2503.01434)
+- VIMS: Visual-Inertial-Magnetic-Sonar SLAM (arxiv 2506.15126)
+
+## Repos GitHub actifs
+- naver/dust3r (7k★): actif, base pipeline lingbot-map
+- Junyi42/monst3r (ICLR 2025): géométrie vidéo dynamique
+- facebookresearch/vggt (CVPR 2025 Best Paper): reconstruction per-frame
+- CUT3R: Continuous 3D Perception, mise à jour mars 2026
+
+## HuggingFace
+- Video-Depth-Anything-Small: depth video temps-réel
+- StereoAdapter: adaptation profondeur stéréo sous-marine
+
+## Signal fort
+VISO (arxiv 2601.01144): pipeline USBL+caméra+IMU pour AUV, pourrait remplacer pure-camera pose estimation dans stage 06_align.