feat(uav): bump min_std from 2. to 4. to reduce detections and run vs…

…s after cluster for auto label

aipipeline/projects/uav/cluster_pipeline.py

@@ -27,7 +27,7 @@
 ENVIRONMENT = os.getenv("ENVIRONMENT") if os.getenv("ENVIRONMENT") else None
 
 
-def process_mission(element):
+def process_mission(element, run_vss=False):
     # Data is in the format
     # <path>,<tator section>,<start image>,<end image>
     # /mnt/UAV/Level-1/trinity-2_20240702T153433_NewBrighton/SONY_DSC-RX1RM2,2024/07/NewBrighton,DSC00100.JPG,DSC00301.JPG
@@ -68,8 +68,13 @@ def process_mission(element):
         "3",
         "--device",
         "cuda:0",
+        "--skip-visualization",
     ]
 
+    if run_vss:
+        vss_url = config_dict["vss"]["url"]
+        args.extend(["--vss-url", vss_url])
+
     container = run_docker(
         image=config_dict["docker"]["sdcat"],
         name=f"sdcat-clu-{mission_name}",
@@ -99,14 +104,18 @@ def run_pipeline(argv=None):
     options = PipelineOptions(beam_args)
     conf_files, config_dict = setup_config(args.config)
 
+    run_vss = False
+    if '--vss' in beam_args:
+        run_vss = True
+
     logger.info("Starting cluster pipeline...")
     with beam.Pipeline(options=options) as p:
         (
             p
             | "Read missions" >> beam.io.ReadFromText(args.missions)
             | "Filter comments" >> beam.Filter(lambda line: not line.startswith("#"))
             | "Create elements" >> beam.Map(lambda line: (line, config_dict))
-            | "Process missions (cluster)" >> beam.Map(process_mission)
+            | "Process missions (cluster)" >> beam.Map(process_mission, run_vss=run_vss)
         )
 
 

aipipeline/projects/uav/config/sdcat.ini

@@ -56,7 +56,7 @@ clahe = False
 block_size = 39
 # The minimum standard deviation of the detection found by the saliency detection to be considered
 # Set to 0 to keep all detections. 8.0 is a good value for general detection, 5.0 is more conservative
-min_std = 2.0
+min_std = 4.0
 # These classes are the ones that showed the best performance in hustvl/yolos-small model for general detection
 ;allowable_classes = person,airplane,boat,bird,kite,surfboard,dog,frisbee,horse,tennis ball,sports ball,animal
 # By setting class_agnostic to True, the model will not use the class information to filter out detections