The final work could be in name of two people?
The answer is yes:
“individuals or teams are invited to enroll in the contest.” – see the Eligibility Criteria from the Rules menu.
Moreover, in this case:
“Each team shall designate a team leader to represent it. The leader will handle all communication with Organizers. The team leader shall coordinate with the other team members and shall be personally responsible for the use and possible distribution of funds provided by Organizers. The Organizers will not accept any claims related to use or distribution of these funds.” – see Contest terms and conditions from the Rules menu.
Please see below the answers to your questions. Thank you for the interest in AICO and looking forward for your solution!
We would like to know which of the libraries listed below are already available/installed on the target system. Additionally, it would be nice to know the exact software version especially for OpenCV and vl_feat.
– OpenCV (release or debug build?)
– Intel IPP, MKL, TBB
At least for OpenCV and vl_feat it would be nice to know if they support Intel IPP, MKL, and TBB. In the case of missing libraries is it possible to provide the binaries (compiled on our system) together with our source code? Which compiler version is installed (GCC, MinGW, Cygwin or Visual Studio)?
Please use only public available libraries (e.g. open source) and deliver your algorithms in compiled form (including the used libraries) for the specified hardware
Considering the requirement document it is clear that the achieved score decreases if CS_Inside is set wrongly, e.g. for too long or at the wrong time. However, it is not clear how the score is determined for the CS_Ahead label. At first, ground-truth conditions are not well defined, i.e. is CS_Ahead set when it is visible the very first time even it is partially occluded or it will be fully covered by an overtaking vehicle in subsequent frames? Secondly, in the graphical scheme it is not apparent that the score decreases when ground-truth is CS_Ahead and the estimated label is CS_Outside. In the third place, considering the situation when the construction site is clearly visible and CS_Ahead is set accordingly. If afterwards an overtaking vehicle causes a full occlusion of the construction site CS_Ahead should be set to CS_Outside since there is no construction site visible anymore. In the requirement document, however, it is stated that CS_Ahead can only be followed by CS_Inside, i.e. toggling of CS_Ahead will be penalized. This means that erroneous ground-truth is provided, since it is absolutely possible that CS_Ahead toggles during approaching the construction site.
Case 1: First time CS_Ahead even when partly occluded:
That algorithm that can deal with partly occluded situation will get most points because of earlier detection.
Case 2: CS_Ahead but estimated label is CS_Outside:
The score do not decrease, but you won’t get points for early detection (CS_Ahead).
Case 3: CS_Ahead but later an occlusion of the construction:
The construction is still there. Setting the state CS_Ahead, CS_Inside and CS_Outside correct requires tracking the states even in case of possible occlusions
Does the testing data set only contain short term construction sites on German highways during daytime following the legal situation or fundamental regulations given by the German regulations called “Regelplan D III” of the “RSA 6/2014”? This is important to know, because the regulations for the nighttime, long term and country road construction sites are totally distinct from “Regelplan D III”, because they do not necessarily contain a mobile warning trailer (sign number 616 StVo) or pylons (sign number 610 StVO).
The test data contain only temporary construction sites with mobile warning trailers at daytime.
Is it possible to get access to measurements of physical quantities of the car like its velocity? Because a good detection algorithm for stationary vehicles on the highway would be achievable if the car’s velocity is known. Such an algorithm is needed to detect the end of a construction site without the optional pylons. In the same context it would be very helpful if we could get access to other measurement channels like radar and steering angle or already calculated bounding boxes of detected objects and road segmentation.
No. Of course, algorithms benefit from an exact ego motion. Other sensors would help to detect objects as well. But we decided to keep the dataset simple and focused on only video data.
Is it allowed to augment the provided training data with publically available images (e.g. data found in the internet)?
Yes, but has to be mentioned in the documentation.
What should we do regarding Sequences where there are no road cones. ( for finding the OUTSIDE of a Construction Site). Where there are no road cones we detect only AHEAD and OUTSIDE, or we need to think at a solution for INSIDE too ? ( how can we know when the Construction Site is over)
Thank you .
Dear Andreea, Alexandru and Sorin,
The INSIDE is not mandatory denoted by cones. The end of a mobil construction site can also be indicated by e.g. construction site vehicle.
Dear Ladies and Gentleman,
we are about to complete our construction site detector for AICO and got one remaining question to the ground truth labels:
Is the training sequence nr. 8 a negativ example?
Because only a few frames at the end of the sequence contain the construction site itself and the full length of the CS isnt captured. Actually the first captured warning sign represents only a optional prewarning sign.
Should we detect the first or the second warning sign as the beginning of the CS? Should we set CS_INSIDE to zero after the last occurence of a pylon after the first captured prewarning sign?
Thanks in advance.
The answer to the question is:
No, it’s a positive example, the construction site is in the middle of the sequence (see picture below)
Ps: The deadline for uploading the solution is 30th of March
thanks for the previous answers you gave to us. They helped us alot.
Usually the requirements state thate the documentation should not exceed 6 pages. To gave a clear describtion of all parts of our algorithm we do need a few more pages. In order to achieve the maximum points for the documentation we would like to know how much our score will decrease in that case.
Thanks in advance.
You must be logged in to post a comment.