shuinb's picture
Publish DSFQ checkpoint, source code and recorded evaluations
79f000b verified
Raw History Blame Contribute Delete
1.38 kB
Quantized-LingBot-Map: DSFQ release artifacts
The DSFQ checkpoint is a post-training quantized derivative of the
LingBot-Map checkpoint from Robbyant and its contributors:
https://huggingface.co/robbyant/lingbot-map
https://github.com/Robbyant/lingbot-map
The upstream project states that it is released under Apache License 2.0.
The original network weights are not trained during this calibration.
The derivative changes the GCT Linear representation through Hadamard
rotation, channel smoothing, learned clipping and signed 4-bit packing.
DINO, geometry heads and other unquantized parameters retain their
documented original/BF16 precision.
The accompanying source follows QuantVGGT's quantization formulation:
https://github.com/wlfeng0509/QuantVGGT
QuantVGGT is distributed under the Apache License 2.0. Native S4 kernels
use NVIDIA CUTLASS, which retains its own BSD-3-Clause license and is
installed separately. Original source notices remain in the source archive.
The included result figures derive from the TUM RGB-D dataset and the
Indoor Travel evaluation sequence. TUM RGB-D attribution:
J. Sturm et al., A Benchmark for the Evaluation of RGB-D SLAM Systems,
IROS 2012. https://cvg.cit.tum.de/data/datasets/rgbd-dataset
Raw image datasets and videos are not included. Their original terms apply.
Source repository:
https://github.com/shuinb/Quantized-LingBot-Map