search

feihuzhang / GANet

GA-Net: Guided Aggregation Net for End-to-end Stereo Matching
MIT License
553 stars 135 forks source link

Prediction/Eval on Middlebury dataset #85

Open skumailraza opened 4 years ago

skumailraza commented 4 years ago

I'm trying to evaluate GANet_deep with the pre-trained weights on the Middlebury-2014 dataset but I'm unable to reproduce disparity_map as shown in your readme. Is there anything different in the transform or preprocessing of the .pfm file while evaluating/predicting on middlebury? e.g scaling the value of the disparity_GT for evaluation or any changes in readPFM() specific to Middlebury images? Attached are bicycle2 images for comparison, you can see that ignoring the heatmap color coding, the disparity still looks much coarser.

bicycle2

With your image: image

fenil25 commented 4 years ago

Did you finetune the model on Middlebury dataset or did you just use the pre-trained model(for 10 epochs) on Sceneflow dataset?

fenil25 commented 4 years ago

Also, I was wondering, how much time did it take for you to run the image through the model. For an 816x1056 image, the prediction time for the network is approximately 13 seconds. I ran the model on Google Colab on GPU.

skumailraza commented 4 years ago

Did you finetune the model on Middlebury dataset or did you just use the pre-trained model(for 10 epochs) on Sceneflow dataset?

Both, I used the pre-trained weights as well as fine tuning on Middlebury dataset, but both produce a coarser disparity map.

Jack47746 commented 2 years ago

Did you finetune the model on Middlebury dataset or did you just use the pre-trained model(for 10 epochs) on Sceneflow dataset?

Both, I used the pre-trained weights as well as fine tuning on Middlebury dataset, but both produce a coarser disparity map.

hi，Did you solve this problem？