search

NVIDIA-AI-IOT / redtail

Perception and AI components for autonomous mobile robotics.
BSD 3-Clause "New" or "Revised" License
1.01k stars 346 forks source link

Dataset used for training #156

Open Diksha-Moolchandani opened 3 years ago

Diksha-Moolchandani commented 3 years ago

What is the dataset used for training the NVsmall, NVTiny, and ResNet models present in the stereoDNN/models folder?

Alexey-Kamenev commented 3 years ago

For training and validation we used KITTI dataset. We used Stereo 2015 benchmark dataset only for evaluation, i.e. we did not train/fine-tune models on that dataset (it's too small anyway, 200 samples AFAIR).

Diksha-Moolchandani commented 3 years ago

What is the significance of input size for these models? The inputs in KITTI 2015 training set are of size (375, 1242). And the input size that is mentioned in ResNet-18_2D is 513x257. How am I suposed to run it? ./nvstereo resnet18_2D 1242 375 stereoDNN/models/ResNet-18_2D/TensorRT/trt_weights.bin left_image_path right_image_path ./disp

or

./nvstereo resnet18_2D 513 257 stereoDNN/models/ResNet-18_2D/TensorRT/trt_weights.bin left_image_path right_image_path ./disp

In any case the ground truth will be 1242x375 and then how will I find the error if I use the second method for inference?