hi,
possibly the answer to this question can be found in the runtime repository’s README.md, but i’d like to ask it anyway:
theoretically, it might be advantageous for a pose estimation algorithm to remember the results of computations it did on an “earlier” image and reuse them when processing a “later” image. is that supported, or does the algorithm have to recompute everything “from scratch”?
cheers.
For each chain of images, you’re allowed to use the images up to and including the current one you are making predictions for. You are expressly not allowed to look ahead, but the others within that chain are fair game.
yes, that’s understood. but my question was whether there is a “legal” venue to use the results of computations that the algorithm did “previously”. that could make a difference in terms of execution time.
for example, if the algorithm, after carefully examining the image, infers that the spaceship is a cube, and not a sphere, can it then use that information at a later point in time ( - i.e. for a later image - ) without having to look at the image again?
I’m not totally sure I understand the distinction you’re drawing between parameters/inferences/etc you may extract from an earlier image to use later and “looking at” the image. The cube/sphere example you gave would be fine though. Does that answer the question?
for now it’s fine i think, i don’t have an algorithm yet anyways. i think i’ll look into the runtime repository and learn the details from there first.
i can reformulate the question in this way: when the algorithm is presented a new image, and its task is to estimate the pose of the spaceship in that image, is the RAM flushed before it starts its calculation, or is the RAM in the same state as it was when the algorithm finished the pose estimation for the previous image?
considering the use case of the algorithm, i think it would totally make sense not to flush the RAM after every pose estimation.
Right, the program you write runs normally and sequentially, memory will not be cleared in between images.
Hi there,
Regarding your question, the ability for a pose estimation algorithm to reuse computations from earlier images can indeed be advantageous in terms of efficiency and speed. This concept is often referred to as “temporal coherence” or “temporal consistency.”
Whether or not the algorithm supports this functionality depends on its design and implementation. Some pose estimation algorithms are designed to take advantage of temporal coherence by incorporating mechanisms to remember and reuse previous computations. This can include techniques such as caching intermediate results or storing relevant information from earlier frames.
I hope this will help you.