jetson inference provides many learning models.
Many models use COCO dataset.
but i don’t find the ped100 image dateset.
Where is the ped100 dataset?
Hi @dovmfflem, I believe that was an internal dataset not for redistribution. Also that model was using DetectNet architecture, which supported a single object class and had lower runtime performance than the multi-class SSD-Mobilenet. Going forward, it is recommended to use the included SSD-Mobilenet/Inception models (SSD-Mobilenet-v2 is now the default object detection model in jetson-inference).
As you pointed out, COCO dataset was used to train the SSD-Mobilenet/Inception models, which include person class.
Thank you reply.
Does SSD-Mobilenet-v2 have a person single class learning model?
The pre-trained SSD-Mobilenet models that come with jetson-inference are trained on the 90-class COCO dataset, so they include other classes too. However, if you change the names of the other classes in the class label file to void, then those classes will be ignored during the detection.
hi @dusty_nv, can you explain that how we can change class names to void ? I didn’t understand this.
I want to ignore some class while object detection.
thanks
Hi @meminar, open your model’s labels text file (for example jetson-inference/data/networks/SSD-Mobilenet-v2/ssd_coco_labels.txt) in an editor. Then change the names of the classes that you want to ignore to void, and save the file. When you run the detectnet program again, it should ignore these classes during detection.
1 Like
it get detection time for void label??
Aside from the post-processing/clustering, using the void label doesn’t decrease the detection time. This is because it is still the same network. You would need to retrain the network on a dataset with just the classes you are interested in, in order to make the network smaller and decrease the detection time.