Noble Logistics — AI Moving Estimate Demo

A web demo that measures the 3D volume of furniture from a single room photo and produces a moving estimate (recommended truck tonnage, crew size, and cost). Built on the LabelAny3D single-image 3D reconstruction pipeline.

1. Environment

pip install -r requirements.txt

GPU environment definitions: envs/labelany3d.yml, envs/sam.yml Setup and external dependencies: docs/INSTALL.md

2. Pipeline

run_single_full_pipeline_parallel.sh runs 7 stages with partial parallelism:

[1] Segmentation (SAM3)            ← sequential (everything depends on it)
      │
      ├─ [2] Depth estimation ─────────────────┐
      └─ [3] Object cropping → [4] Amodal       │  run in parallel
                              → [5] Elevation   │
      ─────────────────────────────────────────┘  wait for all
[6] 3D reconstruction (Amodal3R)   ← sequential
[7] Scene layout alignment         ← sequential

• [1] Segmentation: SAM3 generates instance masks from indoor-furniture keyword prompts (chair, desk, sofa, refrigerator, etc.)
• [2] Depth / [3] Cropping: depth estimation and object cropping run in parallel
• [4] Amodal completion / [5] Elevation: run in parallel on the crops
• [6] 3D reconstruction: reconstructs a 3D mesh from _rgba.png
• [7] Alignment: aligns individual objects into one scene coordinate frame, producing 3dbbox.json

Finally, src/calc_volume.py sums the volume of each 3D bounding box in 3dbbox.json.

3. Performance

Predicted truck tonnage vs. ground truth (GT):

House	GT Truck Size (ton)	Model Predicted (ton)
house_005	1	1
house_010	3 ~ 4	4
house_016	2	2

3D bounding box reconstruction results:

4. Usage

Web demo

cd web_demo
pip install -r requirements.txt
# SAM3 lives in the separate `sam` env, so point the server at it for
# interactive segmentation. Adjust the path to your sam env's python.
SAM_PYTHON=/opt/conda/envs/sam/bin/python uvicorn app:app --host 0.0.0.0 --port 8000

Open http://localhost:8000 → upload a room photo → enter moving details (floor, elevator, distance) → get the estimate.

The web demo supports interactive segmentation: when automatic segmentation is inaccurate, you can add, exclude, or edit instances by clicking before running the pipeline (start → click/tap/select → commit).

Standalone server run

bash run_single_full_pipeline_parallel.sh /abs/path/image.jpg

A sample room photo is bundled at samples/sample_room.jpg, so you can try the pipeline right away:

bash run_single_full_pipeline_parallel.sh "$(pwd)/samples/sample_room.jpg"

⚠️ The standalone server run has no instance-exclusion (interactive selection) feature. Every furniture instance detected by SAM3 is included in 3D reconstruction and volume computation as-is. To exclude specific objects, use the web demo's interactive segmentation.

Key environment variables:

Variable	Default	Description
GPU_IDX	0	GPU index to use
OBJ_REC	amodal3r	3D reconstruction backend
MIN_MASK_AREA	800	Minimum mask area (px) filter
SAM3_PROMPTS	indoor-furniture default list	Segmentation prompts (comma-separated)
SAM3_CONF	0.5	SAM3 confidence threshold
SKIP_SEG	0	If 1, reuse an existing segmentation JSON (SEG_JSON)
SAM_PYTHON	/opt/conda/envs/sam/bin/python	Python interpreter for SAM3

Results are saved under experimental_results/single/val/<scene>/; per-stage timings in timing.txt, final boxes in 3dbbox.json.

5. Citation (LabelAny3D)

This demo is built on the single-image 3D reconstruction pipeline from LabelAny3D. If you use it in your research, please cite:

@inproceedings{yao2025labelany3d,
  title={LabelAny3D: Label Any Object 3D in the Wild},
  author={Jin Yao and Radowan Mahmud Redoy and Sebastian Elbaum and Matthew B. Dwyer and Zezhou Cheng},
  booktitle={Neural Information Processing Systems (NeurIPS)},
  year={2025}
}

• Project page: https://uva-computer-vision-lab.github.io/LabelAny3D/
• Paper: https://openreview.net/pdf?id=Q2fU0JDHuW
• Code: https://github.com/UVA-Computer-Vision-Lab/LabelAny3D