Daniel Seichter
Daniel Seichter
If you are able to run `inference_sample.py` with the samples provided by us, the problem seems to be related to your images. Please check that both images are loaded correctly...
Beyond that, as already mentioned by Mona, we need the dtypes and shapes for both images at this line for further debugging.
The problem is related to your depth image - is not a common depth image with depth values encoded in one channel as yours has three channels. It is more...
In our experiments, SE-add leads to better results as the features get weighted using and Squeeze-and-Excitation operation before fusing them. Depending on the input, the model can decide which features...
Note that the labels in NYUv2 are not perfect and the results may vary a bit. Therefore, we repeated each training several times and used a grid search for the...
The best hyperparameters vary depending on the actual network configuration (network depth, block type, fusion, upsampling, ...). For the default hyperparameters, we picked the values that were best for most...
SceneNetRGBD is a synthetic dataset quite far away from reality in terms of physics and scene layout. IMHO it is only useful for pretraining in order to initialize the weights...
It is difficult to give you a specific hint. With Reset34 backbone 1024 channels should only occur after the context module - 514 channels is quite strange, it should not...
It uses the NYU pre-training model to verify the semantic segmentation effect on the Replica dataset
Looks like that you did not change the number of classes or did not adapted the colormaps. I would recommend to create a new dataset class instead of changing the...
You can map predictions that fall into the last 3 classes to void before computing the metric. Another option would be to adapt the model and to strip the weights...