Troubles with object detection

[hadign20]

I am trying to perform object detection on a custom dataset. I have converted the label format to KITTI and created the dataset. I also changed the size of all images to 1024x1024 and changed every number (1248, 384, ...) in the detectnet_network.prototxt file to 1024. But I have a problem. When I start training the model, the mAP is constantly on 0 which means the network doesn't train.

I don't know what the problem is, but I think it might be with the size of the images. Do I have to change the size of bounding boxes as well, or digit deals with that by itself? What should I do to make this work like it does on KITTI dataset?

[user7077]

Did you use the custom class labels when generating the data set: custom class ?

[hadign20]

@user7077 No I didn't. Is it necessary? Because my classes are different than the ones in the link you mentioned.

[lemhell]

@hadi-ghnd Did you have any success with this problem?

[hadign20]

@lemhell I avoided adjusting class labels on the dataset creation page and waited until around 40 epochs, and it started to show some mAP and precision results.

[sulthanashafi]

I have set epochs upto 30 and got MAP =0 for all value.Also the graph is similar to the one shown above.I have used the custom detect net as mentioned down. customNetworkused.txt

[sulthanashafi]

Please help me to modify the custom net such that it can draw bounding box and coverage.

[sani1486]

@hadi-ghnd how big is your bounding boxes ? did you consider changing the stride ?

[sulthanashafi]

bounding box used is 60x60.Stride =16.My object is really small upto 20x20 pixel.Please suggest to what value i should change its stride value and try.

On 05-Apr-2017, at 2:52 PM, sani1486 notifications@github.com wrote:

@hadi-ghnd https://github.com/hadi-ghnd how big is your bounding boxes ? did you consider changing the stride ?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/NVIDIA/DIGITS/issues/1517#issuecomment-291804084, or mute the thread https://github.com/notifications/unsubscribe-auth/AZZckMlJFciJtRSw717hV2vurXbPuVjGks5rs11NgaJpZM4MiNy4.

[sulthanashafi]

Please may i know which all parameters along with stride to be changed;if taking stride=8;

On 05-Apr-2017, at 2:52 PM, sani1486 notifications@github.com wrote:

@hadi-ghnd https://github.com/hadi-ghnd how big is your bounding boxes ? did you consider changing the stride ?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/NVIDIA/DIGITS/issues/1517#issuecomment-291804084, or mute the thread https://github.com/notifications/unsubscribe-auth/AZZckMlJFciJtRSw717hV2vurXbPuVjGks5rs11NgaJpZM4MiNy4.

[sani1486]

@sulthanashafi You can specify the spacing of the grid squares in the training labels by setting the stride in pixels in the detectnet_groundtruth_param layer. For example:

detectnet_groundtruth_param { stride: 16 scale_cvg: 0.4 gridbox_type: GRIDBOX_MIN min_cvg_len: 20 coverage_type: RECTANGULAR image_size_x: 1024 image_size_y: 512 obj_norm: true crop_bboxes: false } In this layer you can also specify an image training patch size (image_size_x, image_size_y). When these parameters are set, every time an image is fed into DetectNet during training it takes a random crop of this size as input. This can be useful if you have very large images in which the objects you wish to detect are very small. https://devblogs.nvidia.com/parallelforall/detectnet-deep-neural-network-object-detection-digits/

this link might help . and i am not sure what else to change along with stride and crop_bboxes .. i am looking for a solution too . let me know about your developments on this

[sulthanashafi]

Thank you.Have you got the coverage and bounding box.If then may i know whether you used pertrained model and also may you share your log file and prototxt used.

On 05-Apr-2017, at 5:21 PM, sani1486 notifications@github.com wrote:

@sulthanashafi https://github.com/sulthanashafi You can specify the spacing of the grid squares in the training labels by setting the stride in pixels in the detectnet_groundtruth_param layer. For example:

detectnet_groundtruth_param { stride: 16 scale_cvg: 0.4 gridbox_type: GRIDBOX_MIN min_cvg_len: 20 coverage_type: RECTANGULAR image_size_x: 1024 image_size_y: 512 obj_norm: true crop_bboxes: false } In this layer you can also specify an image training patch size (image_size_x, image_size_y). When these parameters are set, every time an image is fed into DetectNet during training it takes a random crop of this size as input. This can be useful if you have very large images in which the objects you wish to detect are very small. https://devblogs.nvidia.com/parallelforall/detectnet-deep-neural-network-object-detection-digits/ <x-msg://8/url> this link might help . and i am not sure what else to change along with stride and crop_bboxes .. i am looking for a solution too . let me know about your developments on this

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/NVIDIA/DIGITS/issues/1517#issuecomment-291837492, or mute the thread https://github.com/notifications/unsubscribe-auth/AZZckEEH9JgWfqEM5MLCU_dRdocDgVtwks5rs4AwgaJpZM4MiNy4.