Update readme for NTU

Signed-off-by: Diogo Luvizon <[email protected]>

INSTALL.md

@@ -21,9 +21,13 @@ Install required python packages before you continue:
 We do not provide public datasets within this software. We only provide
 converted annotation files and some useful scripts for practical purposes.
 
+### MPII
+
 Images from MPII should be manually downloaded and placed
 at `datasets/MPII/images`.
 
+### Human3.6M
+
 Videos from Human3.6M should be manually downloaded and placed
 in `datasets/Human3.6M/S*`, e.g. S1, S2, S3, etc. for each subject.
 After that, extract videos with:
@@ -33,7 +37,22 @@ After that, extract videos with:
 ```
 Python2 is used here due to the dependency on cv2 package.
 
+### PennAction
+
 Video frames from PennAction should be manually downloaded and extracted
 in `datasets/PennAction/frames`. The pose annotations and predicted bounding
 boxes will be automatically downloaded by this software.
 
+### NTU
+
+Video frames from NTU should be also manually extracted.
+A Python [script](datasets/NTU/extract-resize-videos.py) is provided to help in
+this task. Python 2 is required.
+
+Additional pose annotation is provided for NTU, which is used to train the pose
+estimation part for this dataset. It is different from the original Kinect
+poses, since it is a composition of 2D coordinates in RGB frames plus depth.
+This additional annotation can be downloaded
+[here](https://drive.google.com/open?id=1eTJPb8q2XCRK8NEC4h17p17JW2DDNwjG)
+(2GB from Google Drive).
+

README.md

@@ -59,6 +59,16 @@ To reproduce our scores, do:
   python3 exp/pennaction/eval_penn_ar_pe_merge.py output/eval-penn
 ```
 
+### 3D action recognition on NTU
+
+For 3D action recognition, the pose estimation model was trained on mixed
+data from MPII, Human3.6 and NTU, and the full model for action recognition was
+trained and fine-tuned on NTU only.
+To reproduce our scores, do:
+```
+  python3 exp/ntu/eval_ntu_ar_pe_merge.py
+```
+
 
 ## Citing