perf: defer GPU->CPU mask transfer to render thread via LazyFrameObjects

Replace bulk segment_output.to_object_dicts() CPU transfer with lazy
materialization — only the last frame per segment is materialized for
IoU tracking continuity, while render frames use LazyFrameObjects that
defer GPU->CPU copy until the render thread needs them. Reduces CUDA
syncs from 100+ per segment to ~3. Also removes deployment_logs.txt.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

deployment_logs.txt

@@ -1,398 +0,0 @@
-data: {"data":"===== Application Startup at 2026-01-23 22:02:27 =====\n","timestamp":"2026-01-23T22:02:27Z"}
-
-data: {"data":"Set HF_HOME to /tmp/huggingface","timestamp":"2026-01-23T22:02:32.613Z"}
-
-data: {"data":"CUDA_VISIBLE_DEVICES not set. All GPUs should be visible.","timestamp":"2026-01-23T22:02:32.613Z"}
-
-data: {"data":"Startup Diagnostics: Torch version 2.9.1+cu128, CUDA available: True, Device count: 4","timestamp":"2026-01-23T22:02:34.118Z"}
-
-data: {"data":"WARNING ⚠️ user config directory '/home/user/.config/Ultralytics' is not writable, using '/tmp/Ultralytics'. Set YOLO_CONFIG_DIR to override.","timestamp":"2026-01-23T22:02:36.713Z"}
-
-data: {"data":"Creating new Ultralytics Settings v0.0.6 file ✅ ","timestamp":"2026-01-23T22:02:36.721Z"}
-
-data: {"data":"View Ultralytics Settings with 'yolo settings' or at '/tmp/Ultralytics/settings.json'","timestamp":"2026-01-23T22:02:36.721Z"}
-
-data: {"data":"Update Settings with 'yolo settings key=value', i.e. 'yolo settings runs_dir=path/to/dir'. For help see https://docs.ultralytics.com/quickstart/#ultralytics-settings.","timestamp":"2026-01-23T22:02:36.721Z"}
-
-data: {"data":"INFO:     Started server process [1]","timestamp":"2026-01-23T22:02:36.837Z"}
-
-data: {"data":"INFO:     Waiting for application startup.","timestamp":"2026-01-23T22:02:36.837Z"}
-
-data: {"data":"INFO:     Application startup complete.","timestamp":"2026-01-23T22:02:36.837Z"}
-
-data: {"data":"INFO:     Uvicorn running on http://0.0.0.0:7860 (Press CTRL+C to quit)","timestamp":"2026-01-23T22:02:36.837Z"}
-
-data: {"data":"INFO:     10.16.14.243:63556 - \"GET /?logs=container HTTP/1.1\" 307 Temporary Redirect","timestamp":"2026-01-23T22:02:38.715Z"}
-
-data: {"data":"INFO:     10.16.42.137:52271 - \"GET /?logs=container HTTP/1.1\" 307 Temporary Redirect","timestamp":"2026-01-23T22:02:38.764Z"}
-
-data: {"data":"INFO:     10.16.14.243:63556 - \"GET /laser/index.html HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:38.884Z"}
-
-data: {"data":"INFO:     10.16.14.243:63556 - \"GET /laser/js/init.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:38.978Z"}
-
-data: {"data":"INFO:     10.16.42.137:52271 - \"GET /laser/style.css HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:38.979Z"}
-
-data: {"data":"INFO:     10.16.14.243:32424 - \"GET /laser/js/core/config.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:39.056Z"}
-
-data: {"data":"INFO:     10.16.42.137:11947 - \"GET /laser/js/core/state.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:39.061Z"}
-
-data: {"data":"INFO:     10.16.42.137:38479 - \"GET /laser/js/core/utils.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:39.061Z"}
-
-data: {"data":"INFO:     10.16.14.243:21985 - \"GET /laser/js/core/physics.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:39.072Z"}
-
-data: {"data":"INFO:     10.16.42.137:61072 - \"GET /laser/js/core/video.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:39.073Z"}
-
-data: {"data":"INFO:     10.16.14.243:19017 - \"GET /laser/js/ui/logging.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:39.073Z"}
-
-data: {"data":"INFO:     10.16.42.137:8536 - \"GET /laser/js/core/hel.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:39.074Z"}
-
-data: {"data":"INFO:     10.16.14.243:55450 - \"GET /laser/js/ui/cards.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:39.075Z"}
-
-data: {"data":"INFO:     10.16.42.137:7723 - \"GET /laser/js/api/client.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:39.075Z"}
-
-data: {"data":"INFO:     10.16.14.243:61397 - \"GET /laser/js/ui/features.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:39.075Z"}
-
-data: {"data":"INFO:     10.16.14.243:47351 - \"GET /laser/js/core/tracker.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:39.075Z"}
-
-data: {"data":"INFO:     10.16.42.137:28176 - \"GET /laser/js/ui/overlays.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:39.076Z"}
-
-data: {"data":"INFO:     10.16.42.137:33696 - \"GET /laser/js/ui/radar.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:39.076Z"}
-
-data: {"data":"INFO:     10.16.14.243:60830 - \"GET /laser/js/main.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:39.076Z"}
-
-data: {"data":"INFO:     10.16.42.137:13066 - \"GET /laser/js/ui/cursor.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:39.077Z"}
-
-data: {"data":"INFO:     10.16.42.137:52984 - \"GET /laser/js/ui/intel.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:39.078Z"}
-
-data: {"data":"INFO:     10.16.14.243:33644 - \"GET /laser/js/ui/trade.js HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:02:39.080Z"}
-
-data: {"data":"INFO:root:Loading Hugging Face YOLOv8 weights spencercdz/YOLOv8m_defence/yolov8m_defence.pt onto cuda:0","timestamp":"2026-01-23T22:02:53.356Z"}
-
-data: {"data":"INFO:root:GPT Output for First Frame:","timestamp":"2026-01-23T22:03:01.919Z"}
-
-data: {"data":"{'T01': {'id': 'T01', 'distance_m': 300.0, 'direction': \"1 o'clock\", 'description': 'Helicopter in the upper right'}, 'T02': {'id': 'T02', 'distance_m': 350.0, 'direction': \"10 o'clock\", 'description': 'Helicopter in the upper left'}, 'T03': {'id': 'T03', 'distance_m': 280.0, 'direction': \"12 o'clock\", 'description': 'Helicopter in the center'}, 'T04': {'id': 'T04', 'distance_m': 320.0, 'direction': \"11 o'clock\", 'description': 'Helicopter slightly left of center'}}","timestamp":"2026-01-23T22:03:01.919Z"}
-
-data: {"data":"INFO:     10.16.14.243:35547 - \"POST /detect/async HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:01.926Z"}
-
-data: {"data":"INFO:root:Detection queries: ['person', 'car', 'truck', 'motorcycle', 'bicycle', 'bus', 'train', 'airplane']","timestamp":"2026-01-23T22:03:01.941Z"}
-
-data: {"data":"INFO:root:Detected 1 GPUs. Loading models in parallel...","timestamp":"2026-01-23T22:03:01.941Z"}
-
-data: {"data":"INFO:root:Loading Hugging Face YOLOv8 weights spencercdz/YOLOv8m_defence/yolov8m_defence.pt onto cuda:0","timestamp":"2026-01-23T22:03:01.942Z"}
-
-data: {"data":"INFO:     10.16.14.243:35547 - \"GET /detect/first-frame/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:02.031Z"}
-
-data: {"data":"INFO:root:Running GPT estimation for video start (Frame 0)...","timestamp":"2026-01-23T22:03:03.012Z"}
-
-data: {"data":"INFO:     10.16.14.243:37805 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:05.147Z"}
-
-data: {"data":"ERROR:root:Worker failed processing frame","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"Traceback (most recent call last):","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"  File \"/app/inference.py\", line 1048, in flush_batch","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"    queue_out.put((idx, processed, detections), timeout=1.0)","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"  File \"/usr/local/lib/python3.10/queue.py\", line 148, in put","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"    raise Full","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"queue.Full","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"During handling of the above exception, another exception occurred:","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"Traceback (most recent call last):","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"  File \"/app/inference.py\", line 1072, in worker_task","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"    flush_batch()","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"  File \"/app/inference.py\", line 1050, in flush_batch","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"    except Full:","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"NameError: name 'Full' is not defined","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"Exception in thread Thread-4 (worker_task):","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"Traceback (most recent call last):","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"  File \"/app/inference.py\", line 1048, in flush_batch","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"    queue_out.put((idx, processed, detections), timeout=1.0)","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"  File \"/usr/local/lib/python3.10/queue.py\", line 148, in put","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"    raise Full","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"queue.Full","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"During handling of the above exception, another exception occurred:","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"Traceback (most recent call last):","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"  File \"/usr/local/lib/python3.10/threading.py\", line 1016, in _bootstrap_inner","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"    self.run()","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"  File \"/usr/local/lib/python3.10/threading.py\", line 953, in run","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"    self._target(*self._args, **self._kwargs)","timestamp":"2026-01-23T22:03:05.472Z"}
-
-data: {"data":"  File \"/app/inference.py\", line 1072, in worker_task","timestamp":"2026-01-23T22:03:05.473Z"}
-
-data: {"data":"    flush_batch()","timestamp":"2026-01-23T22:03:05.473Z"}
-
-data: {"data":"  File \"/app/inference.py\", line 1050, in flush_batch","timestamp":"2026-01-23T22:03:05.473Z"}
-
-data: {"data":"    except Full:","timestamp":"2026-01-23T22:03:05.473Z"}
-
-data: {"data":"NameError: name 'Full' is not defined","timestamp":"2026-01-23T22:03:05.473Z"}
-
-data: {"data":"INFO:     10.16.14.243:37805 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:08.144Z"}
-
-data: {"data":"ERROR:root:Workers stopped unexpectedly.","timestamp":"2026-01-23T22:03:09.919Z"}
-
-data: {"data":"INFO:     10.16.14.243:37805 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:11.146Z"}
-
-data: {"data":"INFO:     10.16.14.243:37805 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:14.146Z"}
-
-data: {"data":"INFO:     10.16.14.243:37805 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:17.147Z"}
-
-data: {"data":"INFO:     10.16.42.137:24819 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:20.146Z"}
-
-data: {"data":"INFO:     10.16.14.243:44429 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:23.148Z"}
-
-data: {"data":"INFO:     10.16.42.137:53311 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:26.145Z"}
-
-data: {"data":"INFO:     10.16.42.137:53311 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:29.146Z"}
-
-data: {"data":"INFO:     10.16.42.137:53862 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:32.152Z"}
-
-data: {"data":"INFO:     10.16.14.243:19151 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:35.152Z"}
-
-data: {"data":"INFO:     10.16.42.137:55390 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:38.148Z"}
-
-data: {"data":"INFO:     10.16.42.137:55390 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:41.150Z"}
-
-data: {"data":"INFO:     10.16.14.243:57552 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:44.145Z"}
-
-data: {"data":"INFO:     10.16.42.137:40211 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:47.147Z"}
-
-data: {"data":"INFO:     10.16.42.137:40211 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:50.145Z"}
-
-data: {"data":"INFO:     10.16.14.243:65208 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:53.153Z"}
-
-data: {"data":"INFO:     10.16.14.243:65208 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:56.149Z"}
-
-data: {"data":"INFO:     10.16.42.137:40007 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:03:59.150Z"}
-
-data: {"data":"INFO:     10.16.42.137:40007 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:02.144Z"}
-
-data: {"data":"INFO:     10.16.14.243:12635 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:05.147Z"}
-
-data: {"data":"INFO:     10.16.14.243:12584 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:08.151Z"}
-
-data: {"data":"INFO:     10.16.42.137:20413 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:11.149Z"}
-
-data: {"data":"INFO:     10.16.42.137:20413 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:14.150Z"}
-
-data: {"data":"INFO:     10.16.14.243:38842 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:17.148Z"}
-
-data: {"data":"INFO:     10.16.14.243:38842 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:20.150Z"}
-
-data: {"data":"INFO:     10.16.14.243:38842 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:23.145Z"}
-
-data: {"data":"INFO:     10.16.14.243:7310 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:26.151Z"}
-
-data: {"data":"INFO:     10.16.42.137:30457 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:29.152Z"}
-
-data: {"data":"INFO:     10.16.14.243:50162 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:32.148Z"}
-
-data: {"data":"INFO:     10.16.14.243:50162 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:35.147Z"}
-
-data: {"data":"INFO:     10.16.14.243:45255 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:38.152Z"}
-
-data: {"data":"INFO:     10.16.14.243:45255 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:41.145Z"}
-
-data: {"data":"INFO:     10.16.42.137:61679 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:44.150Z"}
-
-data: {"data":"INFO:     10.16.42.137:61679 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:47.148Z"}
-
-data: {"data":"INFO:     10.16.42.137:37407 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:50.156Z"}
-
-data: {"data":"INFO:     10.16.42.137:37407 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:53.155Z"}
-
-data: {"data":"INFO:     10.16.14.243:12888 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:56.146Z"}
-
-data: {"data":"INFO:     10.16.42.137:54774 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:04:59.151Z"}
-
-data: {"data":"INFO:     10.16.14.243:13189 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:02.153Z"}
-
-data: {"data":"INFO:     10.16.42.137:7369 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:05.150Z"}
-
-data: {"data":"INFO:     10.16.14.243:44904 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:08.151Z"}
-
-data: {"data":"INFO:     10.16.14.243:44904 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:11.146Z"}
-
-data: {"data":"INFO:     10.16.14.243:44904 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:14.146Z"}
-
-data: {"data":"INFO:     10.16.42.137:38667 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:17.148Z"}
-
-data: {"data":"INFO:     10.16.14.243:17731 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:20.149Z"}
-
-data: {"data":"INFO:     10.16.14.243:17731 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:23.147Z"}
-
-data: {"data":"INFO:     10.16.42.137:27197 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:26.152Z"}
-
-data: {"data":"INFO:     10.16.42.137:27197 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:29.146Z"}
-
-data: {"data":"INFO:     10.16.14.243:61177 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:32.148Z"}
-
-data: {"data":"INFO:     10.16.14.243:61177 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:35.149Z"}
-
-data: {"data":"INFO:     10.16.14.243:61177 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:38.150Z"}
-
-data: {"data":"INFO:     10.16.14.243:61177 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:41.146Z"}
-
-data: {"data":"INFO:     10.16.14.243:61177 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:44.147Z"}
-
-data: {"data":"INFO:     10.16.14.243:61177 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:47.147Z"}
-
-data: {"data":"INFO:     10.16.14.243:61177 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:50.146Z"}
-
-data: {"data":"INFO:     10.16.14.243:61177 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:53.146Z"}
-
-data: {"data":"INFO:     10.16.14.243:61177 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:56.147Z"}
-
-data: {"data":"INFO:     10.16.42.137:41705 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:05:59.152Z"}
-
-data: {"data":"INFO:     10.16.42.137:41705 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:02.149Z"}
-
-data: {"data":"INFO:     10.16.14.243:51116 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:05.147Z"}
-
-data: {"data":"INFO:     10.16.14.243:51116 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:08.147Z"}
-
-data: {"data":"INFO:     10.16.42.137:59918 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:11.148Z"}
-
-data: {"data":"INFO:     10.16.42.137:37252 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:14.153Z"}
-
-data: {"data":"INFO:     10.16.14.243:49604 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:17.152Z"}
-
-data: {"data":"INFO:     10.16.42.137:24055 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:20.148Z"}
-
-data: {"data":"INFO:     10.16.42.137:24055 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:23.148Z"}
-
-data: {"data":"INFO:     10.16.14.243:24858 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:26.149Z"}
-
-data: {"data":"INFO:     10.16.42.137:65148 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:29.152Z"}
-
-data: {"data":"INFO:     10.16.14.243:29064 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:32.154Z"}
-
-data: {"data":"INFO:     10.16.14.243:29064 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:35.151Z"}
-
-data: {"data":"INFO:     10.16.42.137:44620 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:38.152Z"}
-
-data: {"data":"INFO:     10.16.14.243:46299 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:41.149Z"}
-
-data: {"data":"INFO:     10.16.42.137:8876 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:44.150Z"}
-
-data: {"data":"INFO:     10.16.42.137:8876 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:47.149Z"}
-
-data: {"data":"INFO:     10.16.42.137:8876 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:50.149Z"}
-
-data: {"data":"INFO:     10.16.42.137:8876 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:53.147Z"}
-
-data: {"data":"INFO:     10.16.42.137:14006 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:56.154Z"}
-
-data: {"data":"INFO:     10.16.14.243:55621 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:06:59.153Z"}
-
-data: {"data":"INFO:     10.16.14.243:55621 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:02.151Z"}
-
-data: {"data":"INFO:     10.16.14.243:19848 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:05.159Z"}
-
-data: {"data":"INFO:     10.16.42.137:12980 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:08.151Z"}
-
-data: {"data":"INFO:     10.16.42.137:12980 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:11.152Z"}
-
-data: {"data":"INFO:     10.16.14.243:27976 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:14.155Z"}
-
-data: {"data":"INFO:     10.16.42.137:50949 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:17.154Z"}
-
-data: {"data":"INFO:     10.16.14.243:47930 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:20.155Z"}
-
-data: {"data":"INFO:     10.16.42.137:26741 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:23.154Z"}
-
-data: {"data":"INFO:     10.16.14.243:45298 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:26.150Z"}
-
-data: {"data":"INFO:     10.16.14.243:45298 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:29.151Z"}
-
-data: {"data":"INFO:     10.16.14.243:45298 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:32.153Z"}
-
-data: {"data":"INFO:     10.16.14.243:45298 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:35.148Z"}
-
-data: {"data":"INFO:     10.16.14.243:63025 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:38.159Z"}
-
-data: {"data":"INFO:     10.16.42.137:38082 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:41.153Z"}
-
-data: {"data":"INFO:     10.16.42.137:38082 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:44.149Z"}
-
-data: {"data":"INFO:     10.16.14.243:53673 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:47.152Z"}
-
-data: {"data":"INFO:     10.16.14.243:5029 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:50.414Z"}
-
-data: {"data":"INFO:     10.16.42.137:21004 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:53.415Z"}
-
-data: {"data":"INFO:     10.16.42.137:21004 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:56.415Z"}
-
-data: {"data":"INFO:     10.16.14.243:18558 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:07:59.415Z"}
-
-data: {"data":"INFO:     10.16.42.137:21979 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:08:02.416Z"}
-
-data: {"data":"INFO:     10.16.14.243:4153 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:08:05.414Z"}
-
-data: {"data":"INFO:     10.16.42.137:47440 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:08:08.415Z"}
-
-data: {"data":"INFO:     10.16.42.137:47440 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:08:11.412Z"}
-
-data: {"data":"INFO:     10.16.14.243:51902 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:08:14.416Z"}
-
-data: {"data":"INFO:     10.16.42.137:50678 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:08:17.416Z"}
-
-data: {"data":"INFO:     10.16.42.137:50678 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:08:20.411Z"}
-
-data: {"data":"INFO:     10.16.14.243:24592 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:08:23.410Z"}
-
-data: {"data":"INFO:     10.16.14.243:24592 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:08:26.409Z"}
-
-data: {"data":"INFO:     10.16.14.243:43041 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:08:29.415Z"}
-
-data: {"data":"INFO:     10.16.14.243:13327 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:08:32.422Z"}
-
-data: {"data":"INFO:     10.16.42.137:48743 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:08:35.421Z"}
-
-data: {"data":"INFO:     10.16.14.243:64507 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:08:38.412Z"}
-
-data: {"data":"INFO:     10.16.14.243:29935 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:08:41.493Z"}
-
-data: {"data":"INFO:     10.16.14.243:29935 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:08:44.419Z"}
-
-data: {"data":"INFO:     10.16.14.243:12182 - \"GET /detect/status/720717af953045d5816689f0b22bbecd HTTP/1.1\" 200 OK","timestamp":"2026-01-23T22:08:47.428Z"}
-
-data: {"data":"INFO:     Shutting down","timestamp":"2026-01-23T22:08:55.311Z"}
-
-data: {"data":"INFO:     Waiting for application shutdown.","timestamp":"2026-01-23T22:08:55.411Z"}
-
-data: {"data":"INFO:     Application shutdown complete.","timestamp":"2026-01-23T22:08:55.412Z"}
-
-data: {"data":"INFO:     Finished server process [1]","timestamp":"2026-01-23T22:08:55.412Z"}
-

inference.py

@@ -1642,7 +1642,7 @@ def run_grounded_sam2_tracking(
     from PIL import Image as PILImage
 
     from utils.video import extract_frames_to_jpeg_dir
-    from models.segmenters.grounded_sam2 import MaskDictionary, ObjectInfo, SegmentOutput
+    from models.segmenters.grounded_sam2 import MaskDictionary, ObjectInfo, LazyFrameObjects
 
     active_segmenter = segmenter_name or "gsam2_large"
     logging.info(
@@ -1691,6 +1691,10 @@ def run_grounded_sam2_tracking(
                     break
                 fidx, fobjs = item
                 try:
+                    # Deferred GPU->CPU: materialize lazy objects in render thread
+                    if isinstance(fobjs, LazyFrameObjects):
+                        fobjs = fobjs.materialize()
+
                     if _perf_metrics is not None:
                         _t_r = time.perf_counter()
 
@@ -2064,12 +2068,23 @@ def run_grounded_sam2_tracking(
                     _t_track = time.perf_counter()
 
                 def _feed_segment(seg_frames):
+                    """Fallback for empty/carry-forward segments (already CPU)."""
                     for fidx in sorted(seg_frames.keys()):
                         render_in.put((fidx, seg_frames[fidx]))
 
+                def _feed_segment_gpu(segment_output):
+                    """Feed LazyFrameObjects into render_in (GPU->CPU deferred)."""
+                    # Deduplicate: frame_indices has one entry per (frame, obj)
+                    seen = set()
+                    for fi in segment_output.frame_indices:
+                        if fi not in seen:
+                            seen.add(fi)
+                            render_in.put((fi, LazyFrameObjects(segment_output, fi)))
+
                 tracking_results = segmenter.process_video(
                     frame_dir, frame_names, queries,
                     on_segment=_feed_segment,
+                    on_segment_output=_feed_segment_gpu,
                 )
 
                 if _perf_metrics is not None:
@@ -2197,7 +2212,7 @@ def run_grounded_sam2_tracking(
 
                                 mask_dict = MaskDictionary()
                                 mask_dict.add_new_frame_annotation(
-                                    mask_list=torch.tensor(masks).to(seg.device),
+                                    mask_list=masks,
                                     box_list=(
                                         boxes.clone()
                                         if torch.is_tensor(boxes)
@@ -2325,32 +2340,13 @@ def run_grounded_sam2_tracking(
                             next_seg_idx += 1
                             continue
 
-                        # Bulk CPU transfer: 3 CUDA syncs total (was 100+ per-mask syncs)
-                        segment_results = segment_output.to_object_dicts()
-
-                        # Apply remapping to every frame in this segment
-                        for frame_idx, frame_objects in segment_results.items():
-                            remapped: Dict[int, ObjectInfo] = {}
-                            for local_id, obj_info in frame_objects.items():
-                                global_id = remapping.get(local_id)
-                                if global_id is None:
-                                    continue
-                                remapped[global_id] = ObjectInfo(
-                                    instance_id=global_id,
-                                    mask=obj_info.mask,
-                                    class_name=obj_info.class_name,
-                                    x1=obj_info.x1, y1=obj_info.y1,
-                                    x2=obj_info.x2, y2=obj_info.y2,
-                                )
-                            tracking_results[frame_idx] = remapped
-
-                        if _perf_metrics is not None:
-                            _recon_accum_ms += (time.perf_counter() - _t_rc) * 1000.0
-
-                        # Update running tracker with last frame of this segment
-                        if segment_results:
-                            last_fi = max(segment_results.keys())
-                            last_objs = tracking_results.get(last_fi, {})
+                        # Materialize ONLY the last frame for IoU tracking continuity
+                        last_fi = segment_output.last_frame_idx()
+                        if last_fi is not None:
+                            last_objs = segment_output.frame_to_object_dict(
+                                last_fi, remapping=remapping, to_cpu=True,
+                            )
+                            tracking_results[last_fi] = last_objs
                             sam2_masks = MaskDictionary()
                             sam2_masks.labels = copy.deepcopy(last_objs)
                             if last_objs:
@@ -2364,9 +2360,18 @@ def run_grounded_sam2_tracking(
                                         m.shape[-1] if m.ndim >= 2 else 0
                                     )
 
-                        # Feed reconciled frames to render immediately
-                        for fi in range(start_idx, min(start_idx + step, total_frames)):
-                            render_in.put((fi, tracking_results.get(fi, {})))
+                        if _perf_metrics is not None:
+                            _recon_accum_ms += (time.perf_counter() - _t_rc) * 1000.0
+
+                        # Feed LazyFrameObjects to render — GPU->CPU deferred to render workers
+                        seen_fi: set = set()
+                        for fi in segment_output.frame_indices:
+                            if fi not in seen_fi:
+                                seen_fi.add(fi)
+                                render_in.put((
+                                    fi,
+                                    LazyFrameObjects(segment_output, fi, remapping),
+                                ))
 
                         next_seg_idx += 1
 

models/segmenters/grounded_sam2.py

@@ -37,55 +37,6 @@ class ObjectInfo:
     x2: int = 0
     y2: int = 0
 
-    def update_box(self):
-        """Derive bounding box from mask (GPU-native, minimal sync)."""
-        if self.mask is None:
-            return
-        mask = self.mask
-        if not torch.is_tensor(mask):
-            mask = torch.as_tensor(mask)
-
-        rows = mask.any(dim=1)  # (H,) — which rows have any True
-        cols = mask.any(dim=0)  # (W,) — which cols have any True
-
-        if not rows.any():
-            return
-
-        rows_f = rows.float()
-        cols_f = cols.float()
-        H, W = mask.shape[-2], mask.shape[-1]
-
-        bbox = torch.stack([
-            cols_f.argmax(),
-            rows_f.argmax(),
-            W - 1 - cols_f.flip(0).argmax(),
-            H - 1 - rows_f.flip(0).argmax(),
-        ])
-        x1, y1, x2, y2 = bbox.tolist()
-        self.x1 = int(x1)
-        self.y1 = int(y1)
-        self.x2 = int(x2)
-        self.y2 = int(y2)
-
-    @staticmethod
-    def batch_bbox(masks: torch.Tensor):
-        """Compute bboxes for (N, H, W) bool masks. Returns (N,4) cpu int, (N,) cpu bool."""
-        N, H, W = masks.shape
-        rows = masks.any(dim=2)       # (N, H)
-        cols = masks.any(dim=1)       # (N, W)
-        valid = rows.any(dim=1)       # (N,)
-
-        rows_f = rows.float()
-        cols_f = cols.float()
-
-        y_mins = rows_f.argmax(dim=1)
-        y_maxs = H - 1 - rows_f.flip(1).argmax(dim=1)
-        x_mins = cols_f.argmax(dim=1)
-        x_maxs = W - 1 - cols_f.flip(1).argmax(dim=1)
-
-        bboxes = torch.stack([x_mins, y_mins, x_maxs, y_maxs], dim=1)
-        return bboxes.cpu(), valid.cpu()
-
 
 @dataclass
 class MaskDictionary:
@@ -261,29 +212,89 @@ class SegmentOutput:
     class_names: List[str]       # len == count
     device: str = "cpu"
 
-    def to_object_dicts(self) -> Dict[int, Dict[int, "ObjectInfo"]]:
-        """Bulk CPU transfer + ObjectInfo construction. 3 CUDA syncs total."""
-        if self.masks.numel() == 0:
+    def last_frame_idx(self) -> Optional[int]:
+        return self.frame_indices[-1] if self.frame_indices else None
+
+    def frame_to_object_dict(
+        self,
+        frame_idx: int,
+        remapping: Optional[Dict[int, int]] = None,
+        to_cpu: bool = True,
+    ) -> Dict[int, "ObjectInfo"]:
+        """Materialize a single frame's ObjectInfo dict from GPU buffers.
+
+        Args:
+            frame_idx: The frame index to materialize.
+            remapping: Optional local->global ID mapping.
+            to_cpu: If True, transfer mask/bbox to CPU.
+
+        Returns:
+            ``{obj_id: ObjectInfo}`` for the requested frame.
+        """
+        # Build lazy frame index on first call
+        if not hasattr(self, '_frame_index'):
+            idx: Dict[int, List[int]] = {}
+            for i, fi in enumerate(self.frame_indices):
+                idx.setdefault(fi, []).append(i)
+            self._frame_index = idx
+
+        positions = self._frame_index.get(frame_idx)
+        if not positions:
             return {}
-        masks_cpu = self.masks.cpu()       # sync 1
-        bboxes_cpu = self.bboxes.cpu()     # sync 2
-        valid_cpu = self.valid.cpu()       # sync 3
-        result: Dict[int, Dict[int, ObjectInfo]] = {}
-        for i in range(masks_cpu.shape[0]):
-            fi, oid, cn = self.frame_indices[i], self.obj_ids[i], self.class_names[i]
-            if valid_cpu[i]:
-                x1, y1, x2, y2 = int(bboxes_cpu[i, 0]), int(bboxes_cpu[i, 1]), int(bboxes_cpu[i, 2]), int(bboxes_cpu[i, 3])
+
+        result: Dict[int, ObjectInfo] = {}
+        for i in positions:
+            oid = self.obj_ids[i]
+            cn = self.class_names[i]
+            global_id = remapping.get(oid, oid) if remapping else oid
+            mask = self.masks[i]
+            valid = self.valid[i]
+            if valid:
+                bbox = self.bboxes[i]
+                if to_cpu:
+                    mask = mask.cpu()
+                    x1 = int(bbox[0].item())
+                    y1 = int(bbox[1].item())
+                    x2 = int(bbox[2].item())
+                    y2 = int(bbox[3].item())
+                else:
+                    x1, y1 = int(bbox[0]), int(bbox[1])
+                    x2, y2 = int(bbox[2]), int(bbox[3])
             else:
+                if to_cpu:
+                    mask = mask.cpu()
                 x1 = y1 = x2 = y2 = 0
-            info = ObjectInfo(
-                instance_id=oid, mask=masks_cpu[i],
-                class_name=cn, x1=x1, y1=y1, x2=x2, y2=y2,
+
+            result[global_id] = ObjectInfo(
+                instance_id=global_id,
+                mask=mask,
+                class_name=cn,
+                x1=x1, y1=y1, x2=x2, y2=y2,
             )
-            result.setdefault(fi, {})[oid] = info
         return result
 
-    def last_frame_idx(self) -> Optional[int]:
-        return self.frame_indices[-1] if self.frame_indices else None
+
+# ---------------------------------------------------------------------------
+# Lazy frame objects wrapper (deferred GPU->CPU per-frame)
+# ---------------------------------------------------------------------------
+
+@dataclass
+class LazyFrameObjects:
+    """Lightweight wrapper for deferred GPU->CPU materialization.
+
+    Holds a reference to a GPU-resident ``SegmentOutput`` plus frame index
+    and optional ID remapping.  Call ``materialize()`` to perform the
+    GPU->CPU transfer (intended to run in a render worker thread).
+    """
+    segment_output: SegmentOutput
+    frame_idx: int
+    remapping: Optional[Dict[int, int]] = None
+
+    def materialize(self) -> Dict[int, "ObjectInfo"]:
+        """Transfer one frame's data from GPU to CPU and build ObjectInfo dict."""
+        return self.segment_output.frame_to_object_dict(
+            self.frame_idx, remapping=self.remapping, to_cpu=True,
+        )
 
 
 # ---------------------------------------------------------------------------
@@ -421,6 +432,45 @@ class GroundedSAM2Segmenter(Segmenter):
                 )
         logging.info("num_maskmem changed from %d to %d", old, num_maskmem)
 
+    # -- GPU-resident SAM2 predict (skip numpy conversion) ------------------
+
+    def _predict_masks_gpu(
+        self, input_boxes: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Run SAM2 image predictor keeping masks on GPU (skip numpy conversion).
+
+        Calls SAM2's internal ``_prep_prompts`` + ``_predict`` directly,
+        bypassing the public ``predict()`` which converts to numpy.
+
+        Args:
+            input_boxes: (N, 4) float tensor on device.
+
+        Returns:
+            ``(masks, scores)`` — *masks* is ``(N, H, W)`` bool GPU tensor,
+            *scores* is ``(N,)`` float GPU tensor.
+        """
+        mask_input, unnorm_coords, labels, unnorm_box = (
+            self._image_predictor._prep_prompts(
+                point_coords=None,
+                point_labels=None,
+                box=input_boxes,
+                mask_input=None,
+                normalize_coords=True,
+            )
+        )
+        masks, scores, _ = self._image_predictor._predict(
+            unnorm_coords, labels, unnorm_box, mask_input,
+            multimask_output=False, return_logits=False,
+        )
+        # _predict returns (1, N, ..., H, W); squeeze batch dim
+        masks = masks.squeeze(0)
+        if masks.ndim == 2:
+            masks = masks[None]
+        elif masks.ndim == 4:
+            masks = masks.squeeze(1)
+        scores = scores.squeeze(0).flatten()
+        return masks, scores
+
     # -- Single-frame interface (Segmenter.predict) -------------------------
 
     def predict(
@@ -488,7 +538,7 @@ class GroundedSAM2Segmenter(Segmenter):
         self,
         image: "Image",
         text_prompts: List[str],
-    ) -> Tuple[Optional[np.ndarray], Optional[torch.Tensor], List[str]]:
+    ) -> Tuple[Optional[torch.Tensor], Optional[torch.Tensor], List[str]]:
         """Run GDINO + SAM2 image predictor on a single keyframe.
 
         Args:
@@ -497,9 +547,9 @@ class GroundedSAM2Segmenter(Segmenter):
 
         Returns:
             ``(masks, boxes, labels)`` where *masks* is an ``(N, H, W)``
-            numpy array, *boxes* is an ``(N, 4)`` tensor on device, and
-            *labels* is a list of strings.  Returns ``(None, None, [])``
-            when no objects are detected.
+            bool GPU tensor, *boxes* is an ``(N, 4)`` tensor on device,
+            and *labels* is a list of strings.  Returns
+            ``(None, None, [])`` when no objects are detected.
         """
         self._ensure_models_loaded()
         _pm = getattr(self, '_perf_metrics', None)
@@ -547,12 +597,7 @@ class GroundedSAM2Segmenter(Segmenter):
             _t1 = time.perf_counter()
 
         self._image_predictor.set_image(np.array(image))
-        masks, scores, logits = self._image_predictor.predict(
-            point_coords=None,
-            point_labels=None,
-            box=input_boxes,
-            multimask_output=False,
-        )
+        masks, scores = self._predict_masks_gpu(input_boxes)
 
         if _pm is not None:
             _pl = getattr(self, '_perf_lock', None)
@@ -562,11 +607,6 @@ class GroundedSAM2Segmenter(Segmenter):
             else:
                 _pm["sam_image_total_ms"] += _d
 
-        if masks.ndim == 2:
-            masks = masks[None]
-        elif masks.ndim == 4:
-            masks = masks.squeeze(1)
-
         return masks, input_boxes, det_labels
 
     def propagate_segment(
@@ -579,7 +619,7 @@ class GroundedSAM2Segmenter(Segmenter):
         """Propagate masks for a single segment via SAM2 video predictor.
 
         Returns a GPU-resident ``SegmentOutput`` with zero CUDA syncs.
-        Call ``output.to_object_dicts()`` to materialize CPU ObjectInfo dicts.
+        Call ``output.frame_to_object_dict()`` to materialize per-frame CPU dicts.
         """
         _pm = getattr(self, '_perf_metrics', None)
         if _pm is not None:
@@ -674,6 +714,7 @@ class GroundedSAM2Segmenter(Segmenter):
         frame_names: List[str],
         text_prompts: List[str],
         on_segment: Optional[Callable[[Dict[int, Dict[int, "ObjectInfo"]]], None]] = None,
+        on_segment_output: Optional[Callable[["SegmentOutput"], None]] = None,
     ) -> Dict[int, Dict[int, ObjectInfo]]:
         """Run full Grounded-SAM-2 tracking pipeline on extracted JPEG frames.
 
@@ -801,12 +842,7 @@ class GroundedSAM2Segmenter(Segmenter):
                     _t_si = time.perf_counter()
 
                 self._image_predictor.set_image(np.array(image))
-                masks, scores, logits = self._image_predictor.predict(
-                    point_coords=None,
-                    point_labels=None,
-                    box=input_boxes,
-                    multimask_output=False,
-                )
+                masks, scores = self._predict_masks_gpu(input_boxes)
 
                 if _pm is not None:
                     _pl = getattr(self, '_perf_lock', None)
@@ -816,16 +852,8 @@ class GroundedSAM2Segmenter(Segmenter):
                     else:
                         _pm["sam_image_total_ms"] += _d
 
-                # Normalize mask dims
-                if masks.ndim == 2:
-                    masks = masks[None]
-                    scores = scores[None]
-                    logits = logits[None]
-                elif masks.ndim == 4:
-                    masks = masks.squeeze(1)
-
                 mask_dict.add_new_frame_annotation(
-                    mask_list=torch.tensor(masks).to(device),
+                    mask_list=masks,
                     box_list=input_boxes.clone() if torch.is_tensor(input_boxes) else torch.tensor(input_boxes),
                     label_list=det_labels,
                 )
@@ -863,22 +891,23 @@ class GroundedSAM2Segmenter(Segmenter):
                 segment_output = self.propagate_segment(
                     inference_state, start_idx, mask_dict, step,
                 )
-                segment_results = segment_output.to_object_dicts()
 
-                if segment_results:
-                    all_results.update(segment_results)
-                    if on_segment:
-                        on_segment(segment_results)
-                    last_fi = segment_output.last_frame_idx()
-                    if last_fi is not None:
-                        last_frame_objects = all_results.get(last_fi, {})
-                        sam2_masks = MaskDictionary()
-                        sam2_masks.labels = copy.deepcopy(last_frame_objects)
-                        if last_frame_objects:
-                            first_info = next(iter(last_frame_objects.values()))
-                            if first_info.mask is not None:
-                                sam2_masks.mask_height = first_info.mask.shape[-2] if first_info.mask.ndim >= 2 else 0
-                                sam2_masks.mask_width = first_info.mask.shape[-1] if first_info.mask.ndim >= 2 else 0
+                # GPU-deferred path: only materialize last frame for IoU
+                last_fi = segment_output.last_frame_idx()
+                if last_fi is not None:
+                    last_frame_objects = segment_output.frame_to_object_dict(
+                        last_fi, to_cpu=True,
+                    )
+                    all_results[last_fi] = last_frame_objects
+                    sam2_masks = MaskDictionary()
+                    sam2_masks.labels = copy.deepcopy(last_frame_objects)
+                    if last_frame_objects:
+                        first_info = next(iter(last_frame_objects.values()))
+                        if first_info.mask is not None:
+                            sam2_masks.mask_height = first_info.mask.shape[-2] if first_info.mask.ndim >= 2 else 0
+                            sam2_masks.mask_width = first_info.mask.shape[-1] if first_info.mask.ndim >= 2 else 0
+                if on_segment_output is not None:
+                    on_segment_output(segment_output)
 
         logging.info(
             "Grounded-SAM-2 tracking complete: %d frames, %d tracked objects",