vato007/fast-depth-tf
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
53 Commits 2 Branches 0 Tags
42fcf5554a1813781c61a2c1de0665822e86b024
Commit Graph

53 Commits

This Branch
This Branch
All Branches
Author SHA1 Message Date
Piv
42fcf5554a Update warp test to verify output shape 2021-08-24 21:39:10 +09:30
Piv
c164c9720a Finish Projective Inverse Warp algorithm 2021-08-24 20:13:30 +09:30
Piv
b7917ec465 More warp implementation 2021-08-21 17:32:16 +09:30
Piv
df1ac89a81 Add euler to rotation matrix, grid flattening 2021-08-10 20:39:52 +09:30
Piv
8016f0f945 Add coordinates generation implementation 2021-08-08 22:11:50 +09:30
Piv
ece37843ce Merge branch 'main' into unsupervised 2021-08-08 18:58:39 +09:30
Michael Pivato
a72f04006f Merge branch 'packnet_small' into 'main' 
Add small option to packnet, fix docs and first/final conv layers (prev 32)

See merge request vato007/fast-depth-tf!8
 2021-08-08 09:25:56 +00:00
Michael Pivato
625ecba731 Add small option to packnet, fix docs and first/final conv layers (prev 32) 2021-08-08 09:25:55 +00:00
Michael Pivato
5373dc6b65 Merge branch 'packnet_fixes' into 'main' 
Fix packnet residual block and layers, refactor to support different amount of residual layers

See merge request vato007/fast-depth-tf!7
 2021-08-07 11:31:22 +00:00
Piv
58b8e53986 Fix packnet residual block and layers, refactor to support different amount of residual layers 
I noticed the number of parameters didn't match up to the paper (~128 million)
Fixed this by doing the following:
 - Kernel size of 1 for 3rd conv2d in residual block
 - Use add rather than concat in residual block
 - Fixed add/concat features in decode layers
 - Fixed final layers -> this also allows features_3d == 16 to work
 2021-08-07 21:00:17 +09:30
Piv
cd278e683f Start adding pose warp conversions 2021-08-07 17:18:06 +09:30
Michael Pivato
f56e663fca Merge branch 'unsupervised' into 'main' 
Smooth Loss

See merge request vato007/fast-depth-tf!6
 2021-08-05 08:20:31 +00:00
Michael Pivato
26dda68523 Add Smooth Loss 2021-08-05 08:20:31 +00:00
Piv
5996d6eaf0 Merge branch 'main' into unsupervised 2021-08-05 17:49:48 +09:30
Piv
8be4ce4e6d Add smooth loss 2021-08-05 17:48:44 +09:30
Michael Pivato
e96e6c2c2b Merge branch 'unsupervised' into 'main' 
PoseNet

See merge request vato007/fast-depth-tf!5
 2021-08-04 11:36:30 +00:00
Michael Pivato
5d0731b60f PoseNet 
Adds the nns required for unsupervised learning
 - PoseNet (based on ResNet18 like Monodepth)
 - Wrapper for fast depth to pull out intermediary layers to smooth out gradients during training
 2021-08-04 11:36:30 +00:00
Piv
b95442bb23 Finish off pose net 2021-08-04 20:51:46 +09:30
Piv
a111f89722 Start adding pose decoder 2021-08-03 20:25:19 +09:30
Piv
2372b906df Add resnet18 2021-08-01 10:44:33 +09:30
Michael Pivato
d3a63c6bcd Merge branch 'unsupervised' into 'main' 
Packnet

See merge request vato007/fast-depth-tf!4
 2021-07-31 01:31:25 +00:00
Piv
6514fb0e86 Add sample openvino inference 2021-07-29 20:52:14 +09:30
Piv
3254eef4bf Add compiling packnet model, refactor modules to not duplicate loaders and trainers 2021-07-23 22:41:46 +09:30
Piv
66cbc7faf6 Add pack layer tests, fix unpack_3d layer 2021-07-19 20:37:54 +09:30
Piv
38e7ad069e Refactor load/util, start fixing packnet to support NHWC format 2021-07-19 12:32:56 +09:30
Piv
d8bf493999 More packnet implementation 2021-07-18 19:54:28 +09:30
Piv
603de2bc9f Start adding packnet model 2021-07-18 18:59:25 +09:30
Piv
e372fe33ba Add per-pixel loss functions 2021-07-13 20:32:45 +09:30
Piv
b9457f17fe Add pixel loss functions, move warp 2021-07-13 19:20:24 +09:30
Piv
6f7da21977 Add spatial transformer network sampler 2021-07-12 18:19:59 +09:30
Piv
f501beb6f2 Start implementing unsupervised train loop, add sfmlearner train and utils files for reference 2021-07-05 20:50:12 +09:30
Piv
ba0ba609a3 Add utils from sfm learner, start adding sfm learner loss function to work with keras 2021-06-22 22:03:23 +09:30
Piv
101fe08924 Start adding stubs for unsupervised training 2021-06-16 21:55:13 +09:30
Piv
e5b07fb766 Fix loss function 2021-06-16 21:52:31 +09:30
Michael Pivato
0547509689 Merge branch 'kitti_depth_dataset' into 'main' 
Kitti depth dataset

See merge request vato007/fast-depth-tf!2
 2021-04-22 12:13:48 +00:00
Michael Pivato
070aec6eed Add Kitti depth dataset 
Warning: Using this requires >175gb of disk space (tensorflow will also generate examples that will take up space)
 2021-04-22 12:13:48 +00:00
Piv
02d8cd5810 Remove half-features from dense_depth 2021-04-22 12:30:30 +09:30
Piv
acdb58396c Remove usage of keras 2021-04-14 12:45:01 +09:30
Piv
cf7d2561ec Implement details of dense depth paper 2021-04-14 12:38:51 +09:30
Piv
f598005b73 Add basic coreml and mlkit conversion scripts 2021-03-29 19:08:55 +10:30
Piv
f3fc0f8fbb Add coreml conversion 2021-03-29 18:58:16 +10:30
Michael Pivato
f2a42cca4c Merge branch 'dense-depth' into 'main' 
Dense depth

See merge request vato007/fast-depth-tf!1
 2021-03-29 07:31:42 +00:00
Piv
d88e9d3f12 Add dense-depth and experimental dense-net-nnconv5 models 
Since dense-depth will use half labels by default, the nyu train/eval datasets can be loaded from here at half resolutions for labels
 2021-03-29 17:59:18 +10:30
Piv
870429c3ef Refactor fast-depth 
Addresses the following:
 - Rename nnconv5 block to nnconv5
 - Add skip connections directly to nnconv5 block
 - Allow custom metrics, loss and optimizer (keep defaults that reflect original paper) to train
 - Correctly use nyu evaluation dataset only when no dataset is provided
 2021-03-29 17:57:12 +10:30
Piv
3325ea0c0c Format pep8, include pass shaped to mobilenet 2021-03-25 21:56:50 +10:30
Piv
9449ddef01 Add notebook, gitignore 2021-03-25 21:50:19 +10:30
Piv
78d5aace15 Remove Experimental model 
It didn't perform any better than the regular model
Removing batch normalisation significantly harmed training performance
 2021-03-25 21:28:07 +10:30
Piv
ab7da5acd4 Add documentation and README, use Upsampling2D rather than image Resizing layer 2021-03-24 21:35:25 +10:30
Piv
ac3ab27ddd Add model with no batch normalisation, use actual mobilnet model rather than extracting layers 
Found from DenseDepth, each layer can be set to trainable in the encoder, then the outputs of the
model and the required layers for skip connections can be used directly. Ends up being much cleaner
 2021-03-24 20:00:26 +10:30
Piv
39074f22a7 Add working train and eval functions for nyu_v2 2021-03-21 09:51:45 +10:30