Mapillary Camera Grant Program has launched


#21

I would be interested in the forthcoming bike option to try to record pictures using a gimbal on my bicycle. I already did some tests with a smartphone but a lighter device would be welcome. Does this sound in accordance with this program?

Although my present smartphone is able to take more than one image each 2 seconds, it may encounter overheating (75°C :hushed:):


#22

GoPro Fusion stitching
I apologise, I misspoke on the Fusion. It is not a camera I have tested myself yet. I have edited my earlier post to avoid confusing others.

Server side stitching of Fusion images is not something we have rolled out at scale yet and requires further refinement. Thus client-side stitching is still the only option if you have this camera. It is a good camera but requires patience and a decent machine.

Camera Grant Program
Just to point out, we have been lending cameras for years :stuck_out_tongue_winking_eye:. First for humanitarian scenarios, then as a camera lending program. We have been careful promoting because we wanted to avoid the impression that this was an easy way to get free gear. But I am really happy people are taking to the idea and are appreciative of how it could improve image camera.

Quality consideration
I definitely agree with you on this point. The challenge is judging image quality at scale for many different contributors across the world. We’re applying computer vision to this problem, but we are not there yet. I’d love to hear your thoughts on how you think we can determine this better in the meantime for challenges that we are running. We have a new scoring system coming soon that will factor in the optimal ratio of images to each KM, but it won’t be looking at the contents of an image or the resolution/camera attributes.


#23

This is exactly the kind of scenario the program wants to address. A lot of people (myself included) struggled to take good imagery on a bike with the phone.

I hope to include bike kits in the next wave of camera grants that will go out next month. The delay has been my search for a good mount that enables you to use the GoPro on both a bicycle and a motorbike. I think we’ve found one.


#24

Server side stitching, especially if flow/computational, would be awesome, but can see how this is a longer term effort. Good work thinking about it.

Not saying you just started this programme, but as others pointed out Telenav has a bit more scale. I don’t personally contribute to OSC, but would be really happy to see some consolidation.

I don’t have the details you have on your current efforts, but I was thinking exactly that: ML recog of image that have more detail/shot better/have a wider angle. But then you have the issue of whether people already taking good pictures need to participate in the programme


#25

Ok. I will wait for the bike camera program. My first goal is to do city recording at street level with gimbal. Last summer, I also did a try out of the city but still on concrete road with the phone fixed on a selfy holder (and the holder fixed on the mountain bike, no gimbal). Sample here. Indeed, vibrations induced some horizontal striped blurring. I never tried with mountain bike on mountain paths but I would like to do it.


#26

Is the goal to reduce motion blur with a gimbal ?


#27

No, the gimbal is used to preserve horizontality of device/smartphone (and also reduce zigzag when biking up).

My vibration blur should be related to rolling shutter. I hope that using a GoPro instead of my entry level smartphone will help reducing this.


#28

If you have a gimbal you can put the gimbal+camera on a small monopod and put them in a backpack. Then the body acts as a really great vibration absorber.


#29

You should tell in your profile which camera you use. Maybe then someone can help you. I dare say no more.


#30

Since you’re doing computer vision on all the photos, you already have number of quality statistics for each photo. The photo analysis produces a percentage of how certain the algorithm is about what is has identified. I would say that a higher certainty equals a better quality photo. In addition, you could count the number of unique objects identified per sequence. That is so that sequences of 30 photos per second (frames from video) or letting the camera run while standing still for a traffic light, would not boost your score.


#31

Those are good ideas, but we’d need to make this available in API before we could add it to the scoring system. This would be a good addition though.

We’d also need to consider how this score differs between geographies. A desert road in Australia would have less unique objects per sequence than a main road in Amsterdam. Both would be valued sequences though.

Our new score is not perfect, but it will ensure the 30 frames a second and traffic light scenarios are not scored unfairly.


#32

Nitpicking here, but the new scoring system may still be rewarding the traffic light scenario. Assuming that you calculate the density over the complete sequence. Say I’m having an average of 10m between each image while driving, then letting the camera run while standing still may enhance my overall sequence average to 5m per image (0.2 image/meter).