Yeah, DepthAnything (especially DA3) is also really good. The space is moving so quickly, and I always try to keep an eye on the trending HF depth estimation models :D
We frame the association model as an instance segmentation ranking problem. Given a set of freight item segmentations, the question becomes which freight item is most likely being scanned? Our model needs to analyze the pose of the workers, consider multiple frames and map the whole scene to 3D.
Thanks!! For now, we are focusing on LTL, but improving the put away process in warehousing with better dimensioning data would be a great next use case. thanks for your thoughts!
thanks!! Our wedge is underbilling in LTL trucking with ~10k relevant cross-docking warehouses across the US and Europe. Carriers lose revenue when shippers understate freight dimensions. We're seeing ~$50k/site/month in recoverable revenue from fixing that alone.
And yes, from there we expand to other industries such as fulfillment and manufacturing. Long term, we will be the CV layer for any warehouse running CCTV.
1 + 4) If the bbox fit is accurate, we are below 1.5 inch MAE today. Improving bbox fit accuracy is where most of our effort goes. We're confident this gets to <1 inch at full coverage. The tail is bounded by data and model scale, both of which we're actively closing.
2) Not necessarily. Models like MapAnything/MoGe predict calibration params directly and GeoCalib is good for distortion coefficients. We still calibrate manually on-site, but mainly to validate these models actually hold up in real warehouses and collect our own calibration training data. We are confident the future is calibration-free.
3) carriers lose money every day because shippers understate dimensions and LTL is priced by volume. Every understated shipment is lost revenue. Thats the wedge we sre going after.