Update README.md

README.md

@@ -41,22 +41,8 @@ Global
 
 ### Release Date:
 
-Hugging Face 10/23/2026 via [https://huggingface.co/nvidia/metrictrails](https://huggingface.co/nvidia/metrictrails)
-
-## References(s):
-
-- *MetricTrail: Revisiting the Power of Metric Depth and Local Optimization for
-  Long-Sequence Streaming 3D Reconstruction*.
-- [Repository overview and running instructions](../README.md).
-- [DPVO](https://github.com/princeton-vl/DPVO), the sparse local odometry backbone.
-- [DINOv3](https://github.com/facebookresearch/dinov3) and the
-  [ViT-S/16 LVD-1689M pretrained model](https://huggingface.co/facebook/dinov3-vits16-pretrain-lvd1689m).
-- [MoGe](https://github.com/microsoft/MoGe) and
-  [MoGe-2](https://arxiv.org/abs/2507.02546), the default metric-depth provider.
-- [SALAD](https://github.com/serizba/salad), an optional external place-retrieval
-  tool obtained and run independently of MetricTrail.
-- [ViPE](https://github.com/nv-tlabs/vipe), used for RGB-only shared-focal calibration.
-- [RoMa](https://github.com/Parskatt/RoMa), used for optional revisit-based focal refinement.
+Hugging Face 11/05/2026 via [https://huggingface.co/nvidia/metrictrail](https://huggingface.co/nvidia/metrictrail)
+
 
 ## Model Architecture:
 
@@ -101,52 +87,6 @@ uses 80 patches per frame and 18 recurrent refinements. The frozen DINOv3
 forward pass uses BF16; adapters, odometry, differentiable bundle adjustment,
 and losses use FP32.
 
-<!--
-## Computational Load (internal review only)
-
-Remove this entire comment from public source and the published model card.
-
-Scope: added compute for the documented MetricTrail adaptation run, including
-fusion warmup and joint fine-tuning. The run uses 16 A100 80GB GPUs for
-approximately 7 hours (112 GPU-hours), processing 240,000 clips of 15 frames at
-480 x 640 with 80 patches per frame and 18 recurrent refinements. Third-party
-DPVO and DINOv3 pretraining, evaluation, and discarded exploratory runs are
-excluded. MoGe-2 is not used during adaptation.
-
-**Cumulative Compute (estimated): 1.3 x 10^18 FLOPs.**
-This is an analytical operation estimate, not a hardware-profiler measurement.
-Counting a multiply-add as two FLOPs gives approximately 2.8 x 10^17 FLOPs for
-frozen DINOv3 forward passes, 2.5 x 10^17 for the DPVO convolutional encoders,
-and 6.9 x 10^17 for recurrent-update matrix operations. The encoder/update
-counts approximate forward plus backward as three times forward, with
-adjustments for the frozen warmup. The subtotal is 1.22 x 10^18 FLOPs; the
-reported value includes an approximate allowance for correlation, adapters,
-bundle adjustment, losses, and other operations. Random edge pruning and
-kernel details are not reconstructed. A reasonable sensitivity range is
-0.8 x 10^18 to 2.0 x 10^18 FLOPs, not a validated confidence interval.
-Sources for the operation count: DPVO/dpvo/net.py, DPVO/dpvo/extractor.py,
-DPVO/dpvo/blocks.py, and metrictrail/runtime/dino_fusion.py.
-
-**Estimated Energy for Model Training: 50 kWh.**
-An ORD hardware query on 10/20/2026 confirmed NVIDIA A100-SXM4-80GB with
-configured and default power limits of 400 W. Applying this rating to the
-documented training run, the prescribed calculation is
-16 x 0.400 kW x 0.80 x 7 hours x 1.4 PUE = 50.176 kWh.
-The 80% power fraction and PUE of 1.4 are the internal guidance defaults,
-not measured power or a verified ORD facility PUE.
-
-**Estimated Emissions for Model Training: 0.021 tCO2e (approximately 21 kgCO2e).**
-50.176 kWh x 410.5 gCO2e/kWh x 0.000001 = 0.020597 tCO2e.
-The emissions factor is the guidance's fallback value; a site-specific
-electricity emissions factor was not available. These are estimates for
-the documented adaptation run, not metered energy or emissions.
-
-Methods:
-https://nvidia.sharepoint.com/sites/TrustworthyAI/SitePages/How-to-Assess-Training-Compute.aspx
-https://nvidia.sharepoint.com/:w:/s/TrustworthyAI/Ecb23D6b-kNLv1c6LCdhSiMBjWFz0MKncBaRcryDsfv-vA?e=lIcO9z
-Hardware specifications: https://www.nvidia.com/en-us/data-center/a100/
--->
-
 ## Input:
 
 **Input Type(s):** A temporally ordered RGB image sequence and camera parameters;
@@ -156,9 +96,7 @@ configuration selects the data adapter and processing settings.
 **Input Format(s):** Prepared PNG frames or KITTI NPZ files containing `images`
 as `uint8 [N,3,H,W]`. The repository also supplies Oxford Spires and LaMAria
 inventory adapters. Loaders produce CHW RGB `float32` images in `[0,1]`.
-Arbitrary video containers are not decoded by this CLI. See the
-[input layouts](../README.md#layout-of-external-files) and
-[data adapters](../metrictrail/datasets.py).
+Arbitrary video containers are not decoded by this CLI. 
 
 Optional loop correction consumes a saved local trajectory NPZ, its successful
 JSON sidecar, and an externally generated candidate JSON containing frame pairs
@@ -263,9 +201,7 @@ metadata permits broader versions, but those combinations have not been validate
 No precompiled DPVO CUDA extensions are included. Before tracking, build
 `cuda_corr`, `cuda_ba`, and `lietorch_backends`
 from source in the inference environment, using a compatible C++ compiler,
-the CUDA toolkit including `nvcc`, and Eigen 3.4.0 headers. See
-[environment instructions](../README.md#environment)
-and [package dependencies](../pyproject.toml).
+the CUDA toolkit including `nvcc`, and Eigen 3.4.0 headers. 
 
 **Supported Hardware Microarchitecture Compatibility:**
 
@@ -283,7 +219,7 @@ Run with `PYTHONNOUSERSITE=1`. The primary command is
 `cuda:0`. `--case`, `--resume`, `--preflight`, and
 `--stop-after {calibration,local}` control selection and stage execution.
 Preflight checks external files and relevant runtime imports; it does not
-validate model quality. See [running instructions](../README.md#running).
+validate model quality. 
 
 The main pipeline ends at the local trajectory with `enable_lc: false`.
 Run optional loop correction separately on CPU with
@@ -328,7 +264,7 @@ obtained separately as `dinov3_vits16_pretrain_lvd1689m-08c60483.pth`.
 Its validated SHA256 is
 `08c60483bc63c04f533611e34bf70b120eedb7240f469bc16e9e20bf344b941d`.
 MoGe-2 and optional calibration and RoMa dependencies are also separate
-downloads; see the [external-file layout](../README.md#layout-of-external-files).
+downloads.
 Optional loop-retrieval tools are obtained and run independently to generate
 the candidate files consumed by the loop-correction command.
 
@@ -462,8 +398,7 @@ results use meters; additional indoor/dynamic trajectory results use centimeters
 Dataset averages give sequence-level results equal weight in the reported
 additional protocols. LC denotes the original SALAD/PyPose loop-closure workflow
 used in our evaluations. These results do not establish numerical or benchmark
-equivalence for the new external-candidate NumPy/SciPy postprocessor. See
-[the loop-correction workflow](../docs/loop-closure.md).
+equivalence for the new external-candidate NumPy/SciPy postprocessor. 
 
 | Benchmark / configuration | Average ATE |
 | --- | ---: |