“RF-DETR is both faster and more accurate and truly open source with an Apache 2.0 license”
Misleading marketing statement.
The catch is that for image resolutions >=700x700pixels (most production usecases), the roboflow license is actually PML1.0 instead of Apache2.0
https://github.com/roboflow/rf-detr#license
That may be true for legacy CNNs but very few production use-cases require such a large resolution with DETRs. The latency scales quadratically with the resolution.
Regardless, you can do whatever resolution you want with the Apache 2.0 model. Just change the config at runtime; it was trained to be resolution agnostic.
You are correct that we also released larger models with a larger backbone under a different, non open-source license.
> The catch is that for image resolutions >=700x700pixels (most production usecases)
Citation needed? 2XL looks like you go up to 800x800 pixel inputs. This isn't the dealbreaker you say it is - all pipelines benefit from thoughtful crop and rescaling before going to inference.
See the url in my comment (search for the term rfdetr-2xlarge).
2XL does indeed go up to 800x800 and has PML1.0 license instead of apache 2.0.
Rescaling is fine for some purposes but but not for all. For many domain-specific (often less common and odd dimensioned) objects, downscaling will severely reduce recall. There is a reason that Roboflow slaps a license that is not open source on those specific architectures.
> See the url in my comment (search for the term rfdetr-2xlarge). 2XL does indeed go up to 800x800 and has PML1.0 license instead of apache 2.0.
All of the models, including the Apache 2.0 ones, can be configured to go higher than 800x800. The difference between the ones with the PML license and the Apache 2.0 ones is the backbone, not the resolution.
I'd suggest you read the ICLR paper[1] which shows clearly the difference between the backbones at various latencies in Figure 1.
> For many domain-specific (often less common and odd dimensioned) objects, downscaling will severely reduce recall.
We released an entire paper[2] at Neurips about the long-tail transferability of models across a multitude of domains and benchmarked RF-DETR against that benchmark. The Apache 2.0 model is pareto optimal over the larger PML model at latencies less than the XL size.
(I'm one of the co-founders of Roboflow and worked on RF-DETR and RF100-VL.)
In the case of a universe of finite size, this analogy explains how there can be such a thing as a finite space without there being boundaries provided the space is (slightly) curved.
So this has less to do with an infinitely sized universe and more with the question of “What exists beyond the edge of the universe if it would be finite in size?”
Solar PV takes up a lot of land.
Rich, densely populated countries (like the Netherlands) find the areas outside of the cities to be more valuable as (recreational) nature, for upcoming housing projects, or for agricultural use.
Another (short term) practical reason is that remote areas where the land is cheap usually have a weaker connection to the electrical grid. As erecting vast solar parks takes months while fortifying the grid takes years (and is very expensive), this creates bottlenecks.
Erecting new solar parks in such remote areas is therefore (temporarily) prohibited:
https://www.sunforson.com/dutch-provinces-of-friesland-and-g...
Here you go:
https://lofar-surveys.org/public_hips/LoTSS_DR2_high_hips/
Note that the map is in radio (150MHz) but you can toggle to other ‘maps’ with different wavelengths.
At 150MHz, most emission is synchrotron emission originating either from supernovas and as such is a proxy for star(forming) regions of a galaxy, or originates from (plumes of accelerated particles from) the region around a galaxy’s super massive black hole.
Arecibo and FAST are giant immovables dishes (essentially a parabolic valley covered by some electrically conducting material). The receivers suspended above these dishes are somewhat moveable, but that is about it. That means they can only observe a small region of the sky exactly overhead (this region changes together with the rotation of the earth). VLBI requires multiple telescopes significantly separated from each other in physical space (think thousands of km) to track the same object on the sky during an observation.
Thus VLBI is hard/impossible for giant immovable dishes like FAST :)
> VLBI requires multiple telescopes significantly separated from each other in physical space (think thousands of km) to track the same object on the sky during an observation. Thus VLBI is hard/impossible for giant immovable dishes like FAST. :)
Good point... But the VLBI people I know are "here, hold my beer..." engineers. And astronomers will use Vegemite for thermal paste, if it lowers their Tsys in a tight spot.
If FAST-class dishes are cheap enough, you could build an assortment of them, pointing in different (fixed) directions. The primary azimuth would still train close to zenith, but the hard part is that suspended receiver assembly, the moving bits, a relatively heavy platform (that eventually fell into Arecibo's dish, after decades of minimal budget and maintenance:-( ... Dang, the Canadians have proposed large dirigibles, tethered like weather balloons, for the primary focus assemblies.
I miss that crowd. My first drive was the NRAO VLBA correlator, 23 (!) years ago. I'm out of the loop now.
About the legal basis: "RDS’ [Royal Dutch Shell] reduction obligation ensues from the unwritten standard of care laid down in Book 6 Section 162 Dutch Civil Code, which means that acting in conflict with what is generally accepted according to unwritten law is unlawful. From this standard of care ensues that when determining the Shell group’s corporate policy, RDS must observe the due care exercised in society. The interpretation of the unwritten standard of care calls for an assessment of all circumstances of the case in question."
Excerpt from the court sentence: https://uitspraken.rechtspraak.nl/inziendocument?id=ECLI:NL:...
Incorrect. It comes from an IPCC report:
"In model pathways with no or limited overshoot of 1.5°C, global net anthropogenic CO2 emissions decline by about 45% from 2010 levels by 2030 (40–60% interquartile range), reaching net zero around 2050 (2045–2055 interquartile range)." Excerpt from paragraph 2.3.5.2 of the court sentence: https://uitspraken.rechtspraak.nl/inziendocument?id=ECLI:NL:...
"The assessment culminates in the conclusion that RDS is obliged to reduce the CO2 emissions of the Shell group’s activities by net 45% at end 2030 relative to 2019 through the Shell group’s corporate policy. This reduction obligation relates to the Shell group’s entire energy portfolio and to the aggregate volume of all emissions (Scope 1 through to 3). It is up to RDS to design the reduction obligation, taking account of its current obligations and other relevant circumstances. The reduction obligation is an obligation of result for the activities of the Shell group, with respect to which RDS may be expected to ensure that the CO2 emissions of the Shell group are reduced to this level. This is a significant best-efforts obligation with respect to the business relations of the Shell group, including the end-users, in which context RDS may be expected to take the necessary steps to remove or prevent the serious risks ensuing from the CO2 emissions generated by the business relations, and to use its influence to limit any lasting consequences as much as possible." Excerpt from the court sentence: https://uitspraken.rechtspraak.nl/inziendocument?id=ECLI:NL:...
Misleading marketing statement.
The catch is that for image resolutions >=700x700pixels (most production usecases), the roboflow license is actually PML1.0 instead of Apache2.0 https://github.com/roboflow/rf-detr#license