train model on 1600*1200 resolution dataset to detect 20*20 objects

[lqian]

I have a customized dataset which image is 1600X1200 resolution, and a few kinds of labeled object's resolution is about 20X20. I train detector model with yolo-voc.cfg just modify the classes and filters parameters. The trained model test shows all kind of object has exact classes. Large object with accurate location. But small objects have low recall rate and departure location. Should I increase the width and height and filters parameters to adapt the high resolution image dataset?

[convolutional] size=1 stride=1 pad=1 filters=100 activation=linear [region] anchors = 1.3221, 1.73145, 3.19275, 4.00944, 5.05587, 8.09892, 9.47112, 4.84053, 11.2364, 10.0071 bias_match=1 classes=15 coords=4 num=5 softmax=1 jitter=.3 rescore=1

Region Avg IOU: 0.548691, Class: 0.841120, Obj: 0.557138, No Obj: 0.003176, Avg Recall: 0.606061, count: 33 Region Avg IOU: 0.632646, Class: 0.936380, Obj: 0.631734, No Obj: 0.003543, Avg Recall: 0.709677, count: 31 Region Avg IOU: 0.538804, Class: 0.823894, Obj: 0.562055, No Obj: 0.003856, Avg Recall: 0.631579, count: 38 Region Avg IOU: 0.655477, Class: 0.928076, Obj: 0.586537, No Obj: 0.003680, Avg Recall: 0.806452, count: 31 Region Avg IOU: 0.522283, Class: 0.919344, Obj: 0.501801, No Obj: 0.003539, Avg Recall: 0.566667, count: 30 Region Avg IOU: 0.662957, Class: 0.935198, Obj: 0.593103, No Obj: 0.005123, Avg Recall: 0.717949, count: 39 Region Avg IOU: 0.459496, Class: 0.712431, Obj: 0.213238, No Obj: 0.019586, Avg Recall: 0.457143, count: 35 Region Avg IOU: 0.663684, Class: 0.893058, Obj: 0.635569, No Obj: 0.005632, Avg Recall: 0.769231, count: 39 80200: 7.376122, 7.005815 avg, 0.000001 rate, 19.665792 seconds, 5132800 images Saving weights to backup/itd.backup Saving weights to backup/itd_final.weights

[lqian]

@TheMikeyR @pjreddie what increase the width, height and filters parameter makes CUDA out of memory error. do you have a piece of suggestion?

[yanchao12122]

yolo detector is not good at detecting small object. you can try faster-rcnn.

[mkhegazy]

Hey, can i use the training images to be larger than 1920X1080 or 1700x900 . ?