LingBot-Map Tutorial: GPU-Aware Inference and Point Cloud Export

In this tutorial, we implement an end-to-end streaming 3D reconstruction pipeline with LingBot-Map. We begin by configuring the input source, reconstruction settings, checkpoint selection, and output controls, then probe the available GPU and automatically tune frame limits, camera iterations, scale frames, and KV-cache parameters according to the detected VRAM. We install the repository and its dependencies, download the pretrained checkpoint, preprocess image or video frames, and construct the GCTStream model with streaming attention and long-range trajectory memory. We then perform mixed-precision inference, decode the predicted camera poses and intrinsic parameters, convert depth maps into world-coordinate point clouds, validate the recovered geometry, visualize the reconstructed scene and camera trajectory, and export the results as PLY, NPZ, or GLB artifacts. Copy CodeCopiedUse a different BrowserCFG = { "scene": "courthouse", "image_folder": None, "video_path": None, "fps": 10, "max_frames": None, "stride": None, "checkpoint": "lingbot-map.pt", "image_size": 518, "patch_size": 14, "use_sdpa": True, "mode": "streaming", "num_scale_frames": None, "keyframe_interval": None, "kv_cache_sliding_window": 64, "camera_num_iterations": None, "offload_to_cpu": True, "window_size": 128, "overlap_keyframes": 8, "conf_percentile": 55.0, "pixel_stride": 2, "max_plot_points": 60000, "export_ply": True, "export_glb": False, "launch_viser": False, "run_ablation": False, "seed": 0, } WORK = "/content" REPO = f"{WORK}/lingbot-map" OUT = f"{WORK}/lingbot_out" import os, sys, subprocess, glob, json, time, math, shutil, textwrap os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True") os.makedirs(OUT, exist_ok=True) def sh(cmd, check=True): p = subprocess.run(cmd, shell=True, capture_output=True, text=True) if p.returncode != 0 and check: print(p.stdout[-3000:]); print(p.stderr[-3000:]) raise RuntimeError(f"command failed: {cmd}") return p def probe_gpu(): try: q = sh("nvidia-smi --query-gpu=name,memory.total --format=csv,noheader", check=False) if q.returncode != 0 or not q.stdout.strip(): return None, 0.0 name, mem = [x.strip() for x in q.stdout.strip().split("\n")[0].split(",")] return name, float(mem.split()[0]) / 1024.0 except Exception: return None, 0.0 GPU_NAME, VRAM_GB = probe_gpu() print("=" * 78) print(f"GPU : {GPU_NAME or 'NONE — enable Runtime > Change runtime type > GPU'}") print(f"VRAM : {VRAM_GB:.1f} GB") try: import psutil print(f"System RAM : {psutil.virtual_memory().total/2**30:.1f} GB") except Exception: pass print(f"Free disk : {shutil.disk_usage(WORK).free/2**30:.1f} GB (checkpoint is 4.6 GB)") print("=" * 78) if GPU_NAME is None: raise SystemExit("This model needs a CUDA GPU. Switch the Colab runtime to GPU.") if VRAM_GB < 18: AUTO = dict(max_frames=48, num_scale_frames=4, camera_num_iterations=2, kvsw=48) tier = "small (<18 GB)" elif VRAM_GB < 30: AUTO = dict(max_frames=96, num_scale_frames=8, camera_num_iterations=4, kvsw=64) tier = "medium (18-30 GB)" else: AUTO = dict(max_frames=240, num_scale_frames=8, camera_num_iterations=4, kvsw=64) tier = "large (30+ GB)" if CFG["max_frames"] is None: CFG["max_frames"] = AUTO["max_frames"] if CFG["num_scale_frames"] is None: CFG["num_scale_frames"] = AUTO["num_scale_frames"] if CFG["camera_num_iterations"] is None: CFG["camera_num_iterations"] = AUTO["camera_num_iterations"] if VRAM_GB < 18: CFG["kv_cache_sliding_window"] = AUTO["kvsw"] print(f"Auto-tuned for {tier}: max_frames={CFG['max_frames']}, " f"scale_frames={CFG['num_scale_frames']}, " f"cam_iters={CFG['camera_num_iterations']}, " f"kv_window={CFG['kv_cache_sliding_win