KITTI DetectNet example: not detecting pedestrians

[DarylWM]

I've been trying to use the KITTI dataset and the DetectNet example to detect pedestrians, but after 100 epochs (batch size 8) the mAP remains zero. I've set custom classes to: dontcare,pedestrian

I haven't changed anything else from the example. For example, I haven't set the transformations in the augmentation parameters to zero. Should I need to?

[JorgeVilchis]

Did make sure to set subtract mean to none when creating the training job in DIGITS ?

[DarylWM]

Yes, "Subtract Mean" is set to None.

Zero mAP even after 150 epochs.

[DarylWM]

It's interesting that mAP increases above zero at epoch 3 when it's learning to detect cyclists, where custom classes is: dontcare,Cyclist

[gheinrich]

If you don't have https://github.com/NVIDIA/DIGITS/commit/8a9b50b055ffd1ef24c898fc74c01e7026003ba3#diff-1771beb4c58b8620538518e01e546321 you need to lower case your custom class names.

[DarylWM]

Thanks @gheinrich, I have that change. My problem is I can't get the model to detect pedestrians; cars and cyclists are detected, but not pedestrians.

[eweill]

I am also trying to accomplish the same task and have tried many different configurations of data set creation (for object detection) as well as modeling styles (different learning rate strategies, etc.). I am having the same problem where I can't get the model to detect pedestrians at all but cars get detected perfectly. No matter if I use a single class pedestrian detector or a 2-class car-pedestrian detector, the pedestrians are never detected.

[DarylWM]

I tried turning off the augmentation parameters just to see if that might help, but it didn't. Any ideas what the difference is between cyclists (which can be detected) and pedestrians (which can't)?

Pedestrians: 4299 in the training set, 188 in the validation set Cyclists: 1445 in the training set, 182 in the validation set

Could it be that the variation of pedestrians is greater than for cyclists?

[s5919]

Have you got any new results?

I am working at the same problem: Pedestrian Detection with DetectNet. At first I thought the KITTI database hasn't got enough or bad pedestrian examples. Therefore I created my own dataset, which is labeled like the KITTI dataset. Unfortunately mAP stays at zero during the training.

Is it possible, that the network is not able to detect pedestrians?

[alpop]

I've had same problem struggling hard with pedestrian detection: map, precision and recall all were zeros even at epoch 300...

Luckily today it started learning after I had taken (KITTY training) images 2000-3499 for training and 5000-5499 for validation and changed LR to 0.00001. No changes in the net.

At epoch 9 map/precision/recall started to increase slowly from zero to around 10 at epoch 30. At this stage I checked the model performance using held-out images and - wow!!! - it started detect pedestrians. I'm planning to play with LR, train/val splits, data augmentation, decay to improve performance.

[DarylWM]

Thanks for the tip @alpop, I'll give that a try.

By the way, for the sake of comparison, I tried Darknet on the pedestrians in the KITTI data set.

[JorgeVilchis]

Nice job ! Do you think the performance in terms of localization / classification is better using DarkNet ?

On Jan 8, 2017, at 11:10 PM, Daryl Wilding-McBride notifications@github.com wrote:

Thanks for the tip @alpop, I'll give that a try.

By the way, for the sake of comparison, I tried Darknet on the pedestrians in the Kitti data set.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

[ShervinAr]

@DarylWM did you manage to get pedestrians detected?

[DarylWM]

Not yet with DetectNet.

On 15 Jan 2017, at 11:37 PM, ShervinAr notifications@github.com wrote:

@DarylWM did you manage to get pedestrians detected?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

[alpop]

Great job, nice picture! What it should look like with DectectNet on held-out image with map around 20:

[ShervinAr]

@alpop could you please post also the original image without bounding boxes so that I could test it with my trained net?

[sulth]

Pedestrian for single class and two class not getting learned.Please someone give me some tips.Also is the darknet approach good and accurate than detectnet?

[sulth]

@DarylWM have you got detectnet multiclass detection run.May i also know if darknet better than detectnet