dronescapes / README.md
Meehai's picture
camera matrics and camera normals
ee41033
|
raw
history blame
21.2 kB
# Dronescapes dataset
![Logo](logo.png)
As introduced in our ICCV 2023 workshop paper: [link](https://openaccess.thecvf.com/content/ICCV2023W/LIMIT/papers/Marcu_Self-Supervised_Hypergraphs_for_Learning_Multiple_World_Interpretations_ICCVW_2023_paper.pdf)
# 1. Downloading the data
## Option 1. Download the pre-processed dataset from HuggingFace repository
```
git lfs install # Make sure you have git-lfs installed (https://git-lfs.com)
git clone https://huggingface.co/datasets/Meehai/dronescapes
```
Note: the dataset has about 300GB, so it may take a while to clone it.
<details>
<summary> <b> Option 2. Generating the dataset from raw videos and basic labels </b>.</summary>
Recommended if you intend on understanding how the dataset was created or add new videos or representations.
### 1.2.1 Raw videos
Follow the commands in each directory under `raw_data/videos/*/commands.txt` if you want to start from the 4K videos.
If you only want the 540p videos as used in the paper, they are already provided in the `raw_data/videos/*` directories.
### 1.2.2 Semantic segmentation labels (human annotated)
These were human annotated and then propagated using [segprop](https://github.com/vlicaret/segprop).
```bash
cd raw_data/
tar -xzvf segprop_npz_540.tar.gz
```
### 1.2.3 Generate the rest of the representations
We use the [video-representations-extractor](https://gitlab.com/meehai/video-representations-extractor) to generate
the rest of the labels using pre-traing networks or algoritms.
```
VRE_DEVICE=cuda CUDA_VISIBLE_DEVICES=0 vre raw_data/videos/atanasie_DJI_0652_full/atanasie_DJI_0652_full_540p.mp4 -o raw_data/npz_540p/atanasie_DJI_0652_full/ --cfg_path scripts/cfg.yaml --batch_size 3 --n_threads_data_storer 4 --output_dir_exist_mode overwrite --representations rgb "opticalflow_rife" "depth_dpt" "edges_dexined" "semantic_mask2former_swin_mapillary"
VRE_DEVICE=cuda CUDA_VISIBLE_DEVICES=1 vre raw_data/videos/barsana_DJI_0500_0501_combined_sliced_2700_14700/barsana_DJI_0500_0501_combined_sliced_2700_14700_540p.mp4 -o raw_data/npz_540p/barsana_DJI_0500_0501_combined_sliced_2700_14700/ --cfg_path scripts/cfg.yaml --batch_size 3 --n_threads_data_storer 4 --output_dir_exist_mode overwrite --representations rgb "opticalflow_rife" "depth_dpt" "edges_dexined" "semantic_mask2former_swin_mapillary"
VRE_DEVICE=cuda CUDA_VISIBLE_DEVICES=2 vre raw_data/videos/comana_DJI_0881_full/comana_DJI_0881_full_540p.mp4 -o raw_data/npz_540p/comana_DJI_0881_full/ --cfg_path scripts/cfg.yaml --batch_size 3 --n_threads_data_storer 4 --output_dir_exist_mode overwrite --representations rgb "opticalflow_rife" "depth_dpt" "edges_dexined" "semantic_mask2former_swin_mapillary"
VRE_DEVICE=cuda CUDA_VISIBLE_DEVICES=3 vre raw_data/videos/gradistei_DJI_0787_0788_0789_combined_sliced_3510_13110/gradistei_DJI_0787_0788_0789_combined_sliced_3510_13110_540p.mp4 -o raw_data/npz_540p/gradistei_DJI_0787_0788_0789_combined_sliced_3510_13110/ --cfg_path scripts/cfg.yaml --batch_size 3 --n_threads_data_storer 4 --output_dir_exist_mode overwrite --representations rgb "opticalflow_rife" "depth_dpt" "edges_dexined" "semantic_mask2former_swin_mapillary"
VRE_DEVICE=cuda CUDA_VISIBLE_DEVICES=4 vre raw_data/videos/herculane_DJI_0021_full/herculane_DJI_0021_full_540p.mp4 -o raw_data/npz_540p/herculane_DJI_0021_full/ --cfg_path scripts/cfg.yaml --batch_size 3 --n_threads_data_storer 4 --output_dir_exist_mode overwrite --representations rgb "opticalflow_rife" "depth_dpt" "edges_dexined" "semantic_mask2former_swin_mapillary"
VRE_DEVICE=cuda CUDA_VISIBLE_DEVICES=5 vre raw_data/videos/jupiter_DJI_0703_0704_0705_combined_sliced_10650_21715/jupiter_DJI_0703_0704_0705_combined_sliced_10650_21715_540p.mp4 -o raw_data/npz_540p/jupiter_DJI_0703_0704_0705_combined_sliced_10650_21715/ --cfg_path scripts/cfg.yaml --batch_size 3 --n_threads_data_storer 4 --output_dir_exist_mode overwrite --representations rgb "opticalflow_rife" "depth_dpt" "edges_dexined" "semantic_mask2former_swin_mapillary"
VRE_DEVICE=cuda CUDA_VISIBLE_DEVICES=6 vre raw_data/videos/norway_210821_DJI_0015_full/norway_210821_DJI_0015_full_540p.mp4 -o raw_data/npz_540p/norway_210821_DJI_0015_full/ --cfg_path scripts/cfg.yaml --batch_size 3 --n_threads_data_storer 4 --output_dir_exist_mode overwrite --representations rgb "opticalflow_rife" "depth_dpt" "edges_dexined" "semantic_mask2former_swin_mapillary"
VRE_DEVICE=cuda CUDA_VISIBLE_DEVICES=7 vre raw_data/videos/olanesti_DJI_0416_full/olanesti_DJI_0416_full_540p.mp4 -o raw_data/npz_540p/olanesti_DJI_0416_full/ --cfg_path scripts/cfg.yaml --batch_size 3 --n_threads_data_storer 4 --output_dir_exist_mode overwrite --representations rgb "opticalflow_rife" "depth_dpt" "edges_dexined" "semantic_mask2former_swin_mapillary"
VRE_DEVICE=cuda CUDA_VISIBLE_DEVICES=0 vre raw_data/videos/petrova_DJI_0525_0526_combined_sliced_2850_11850/petrova_DJI_0525_0526_combined_sliced_2850_11850_540p.mp4 -o raw_data/npz_540p/petrova_DJI_0525_0526_combined_sliced_2850_11850/ --cfg_path scripts/cfg.yaml --batch_size 3 --n_threads_data_storer 4 --output_dir_exist_mode overwrite --representations rgb "opticalflow_rife" "depth_dpt" "edges_dexined" "semantic_mask2former_swin_mapillary"
VRE_DEVICE=cuda CUDA_VISIBLE_DEVICES=1 vre raw_data/videos/slanic_DJI_0956_0957_combined_sliced_780_9780/slanic_DJI_0956_0957_combined_sliced_780_9780_540p.mp4 -o raw_data/npz_540p/slanic_DJI_0956_0957_combined_sliced_780_9780/ --cfg_path scripts/cfg.yaml --batch_size 3 --n_threads_data_storer 4 --output_dir_exist_mode overwrite --representations rgb "opticalflow_rife" "depth_dpt" "edges_dexined" "semantic_mask2former_swin_mapillary"
```
### 1.2.4 Convert Mask2Former from Mapillary classes to segprop8 classes
Since we are using pre-trained Mask2Former which has either mapillary or COCO panoptic classes, we need to convert them to dronescapes-compatible (8) classes.
To do this, we use the `scripts/convert_m2f_to_dronescapes.py` script:
```
python scripts/convert_m2f_to_dronescapes.py in_dir out_dir mapillary/coco [--overwrite]
```
```
python scripts/convert_m2f_to_dronescapes.py raw_data/npz_540p/atanasie_DJI_0652_full/semantic_mask2former_swin_mapillary raw_data/npz_540p/atanasie_DJI_0652_full/semantic_mask2former_swin_mapillary_converted mapillary
python scripts/convert_m2f_to_dronescapes.py raw_data/npz_540p/barsana_DJI_0500_0501_combined_sliced_2700_14700/semantic_mask2former_swin_mapillary raw_data/npz_540p/barsana_DJI_0500_0501_combined_sliced_2700_14700/semantic_mask2former_swin_mapillary_converted mapillary
python scripts/convert_m2f_to_dronescapes.py raw_data/npz_540p/comana_DJI_0881_full/semantic_mask2former_swin_mapillary raw_data/npz_540p/comana_DJI_0881_full/semantic_mask2former_swin_mapillary_converted mapillary
python scripts/convert_m2f_to_dronescapes.py raw_data/npz_540p/gradistei_DJI_0787_0788_0789_combined_sliced_3510_13110/semantic_mask2former_swin_mapillary raw_data/npz_540p/gradistei_DJI_0787_0788_0789_combined_sliced_3510_13110/semantic_mask2former_swin_mapillary_converted mapillary
python scripts/convert_m2f_to_dronescapes.py raw_data/npz_540p/herculane_DJI_0021_full/semantic_mask2former_swin_mapillary raw_data/npz_540p/herculane_DJI_0021_full/semantic_mask2former_swin_mapillary_converted mapillary
python scripts/convert_m2f_to_dronescapes.py raw_data/npz_540p/jupiter_DJI_0703_0704_0705_combined_sliced_10650_21715/semantic_mask2former_swin_mapillary raw_data/npz_540p/jupiter_DJI_0703_0704_0705_combined_sliced_10650_21715/semantic_mask2former_swin_mapillary_converted mapillary
python scripts/convert_m2f_to_dronescapes.py raw_data/npz_540p/norway_210821_DJI_0015_full/semantic_mask2former_swin_mapillary raw_data/npz_540p/norway_210821_DJI_0015_full/semantic_mask2former_swin_mapillary_converted mapillary
python scripts/convert_m2f_to_dronescapes.py raw_data/npz_540p/olanesti_DJI_0416_full/semantic_mask2former_swin_mapillary raw_data/npz_540p/olanesti_DJI_0416_full/semantic_mask2former_swin_mapillary_converted mapillary
python scripts/convert_m2f_to_dronescapes.py raw_data/npz_540p/petrova_DJI_0525_0526_combined_sliced_2850_11850/semantic_mask2former_swin_mapillary raw_data/npz_540p/petrova_DJI_0525_0526_combined_sliced_2850_11850/semantic_mask2former_swin_mapillary_converted mapillary
python scripts/convert_m2f_to_dronescapes.py raw_data/npz_540p/slanic_DJI_0956_0957_combined_sliced_780_9780/semantic_mask2former_swin_mapillary raw_data/npz_540p/slanic_DJI_0956_0957_combined_sliced_780_9780/semantic_mask2former_swin_mapillary_converted mapillary
```
### 1.2.5 Check counts for consistency
Run: `bash scripts/count_npz.sh raw_data/npz_540p`. At this point it should return:
| scene | rgb | depth_dpt | depth_sfm_manual20.. | edges_dexined | normals_sfm_manual.. | opticalflow_rife | semantic_mask2form.. | semantic_segprop8 |
|:----------|------:|------------:|-----------------------:|----------------:|-----------------------:|-------------------:|-----------------------:|--------------------:|
| atanasie | 9021 | 9021 | 9020 | 9021 | 9020 | 9021 | 9021 | 9001 |
| barsana | 12001 | 12001 | 12001 | 12001 | 12001 | 12000 | 12001 | 1573 |
| comana | 9022 | 9022 | 0 | 9022 | 0 | 9022 | 9022 | 1210 |
| gradistei | 9601 | 9601 | 9600 | 9601 | 9600 | 9600 | 9601 | 1210 |
| herculane | 9022 | 9022 | 9021 | 9022 | 9021 | 9022 | 9022 | 1210 |
| jupiter | 11066 | 11066 | 11065 | 11066 | 11065 | 11066 | 11066 | 1452 |
| norway | 2983 | 2983 | 0 | 2983 | 0 | 2983 | 2983 | 2941 |
| olanesti | 9022 | 9022 | 9021 | 9022 | 9021 | 9022 | 9022 | 1210 |
| petrova | 9001 | 9001 | 9001 | 9001 | 9001 | 9000 | 9001 | 1210 |
| slanic | 9001 | 9001 | 9001 | 9001 | 9001 | 9000 | 9001 | 9001 |
### 1.2.6. Split intro train, validation, semisupervised and train
We include 8 splits: 4 using only GT annotated semantic data and 4 using all available data (i.e. segproped between
annotated data). The indexes are taken from `txt_files/*`, i.e. `txt_files/manually_adnotated_files/test_files_116.txt`
refers to the fact that the (unseen at train time) test set (norway + petrova + barsana) contains 116 manually
annotated semantic files. We include all representations from above, not just semantic for all possible splits.
Adding new representations is as simple as running VRE on the 540p mp4 file
```
python scripts/symlinks_from_txt_list.py raw_data/npz_540p/ --txt_file txt_files/annotated_and_segprop/train_files_11664.txt -o data/train_set --overwrite
python scripts/symlinks_from_txt_list.py raw_data/npz_540p/ --txt_file txt_files/annotated_and_segprop/val_files_605.txt -o data/validation_set --overwrite
python scripts/symlinks_from_txt_list.py raw_data/npz_540p/ --txt_file txt_files/annotated_and_segprop/semisup_files_11299.txt -o data/semisupervised_set --overwrite
python scripts/symlinks_from_txt_list.py raw_data/npz_540p/ --txt_file txt_files/annotated_and_segprop/test_files_5603.txt -o data/test_set --overwrite
python scripts/symlinks_from_txt_list.py raw_data/npz_540p/ --txt_file txt_files/manually_annotated_files/train_files_218.txt -o data/train_set_annotated_only --overwrite
python scripts/symlinks_from_txt_list.py raw_data/npz_540p/ --txt_file txt_files/manually_annotated_files/val_files_15.txt -o data/validation_set_annotated_only --overwrite
python scripts/symlinks_from_txt_list.py raw_data/npz_540p/ --txt_file txt_files/manually_annotated_files/semisup_files_207.txt -o data/semisupervised_set_annotated_nly --overwrite
python scripts/symlinks_from_txt_list.py raw_data/npz_540p/ --txt_file txt_files/manually_annotated_files/test_files_116.txt -o data/test_set_annotated_nly --overwrite
```
Note: `add --copy_files` if you want to make copies instead of using symlinks.
Upon calling this, you should be able to see something like this:
```
user> ls data/*
data/semisupervised_set:
depth_dpt edges_dexined opticalflow_rife semantic_mask2former_swin_mapillary_converted
depth_sfm_manual202204 normals_sfm_manual202204 rgb semantic_segprop8
data/semisupervised_set_annotated_nly:
depth_dpt edges_dexined opticalflow_rife semantic_mask2former_swin_mapillary_converted
depth_sfm_manual202204 normals_sfm_manual202204 rgb semantic_segprop8
data/test_set:
depth_dpt edges_dexined opticalflow_rife semantic_mask2former_swin_mapillary_converted
depth_sfm_manual202204 normals_sfm_manual202204 rgb semantic_segprop8
data/test_set_annotated_nly:
depth_dpt edges_dexined opticalflow_rife semantic_mask2former_swin_mapillary_converted
depth_sfm_manual202204 normals_sfm_manual202204 rgb semantic_segprop8
data/train_set:
depth_dpt edges_dexined opticalflow_rife semantic_mask2former_swin_mapillary_converted
depth_sfm_manual202204 normals_sfm_manual202204 rgb semantic_segprop8
data/train_set_annotated_only:
depth_dpt edges_dexined opticalflow_rife semantic_mask2former_swin_mapillary_converted
depth_sfm_manual202204 normals_sfm_manual202204 rgb semantic_segprop8
data/validation_set:
depth_dpt edges_dexined opticalflow_rife semantic_mask2former_swin_mapillary_converted
depth_sfm_manual202204 normals_sfm_manual202204 rgb semantic_segprop8
data/validation_set_annotated_only:
depth_dpt edges_dexined opticalflow_rife semantic_mask2former_swin_mapillary_converted
depth_sfm_manual202204 normals_sfm_manual202204 rgb semantic_segprop8
```
### 1.2.7 Convert Camera Normals to World Normals
This is an optional step, but for some use cases, it may be better to use world normals instead of camera normals, which
are provided by default in `normals_sfm_manual202204`. To convert, we provide camera rotation matrices in
`raw_data/camera_matrics.tar.gz` for all 8 scenes that also have SfM.
In order to convert, use this function (for each npz file):
```
def convert_camera_to_world(normals: np.ndarray, rotation_matrix: np.ndarray) -> np.ndarray:
normals = (normals.copy() - 0.5) * 2 # [-1:1] -> [0:1]
camera_normals = camera_normals @ np.linalg.inv(rotation_matrix)
camera_normals = (camera_normals / 2) + 0.5 # [0:1] => [-1:1]
return np.clip(camera_normals, 0.0, 1.0)
```
</details>
## 2. Using the data
As per the split from the paper:
![Split](split.png)
The data is in `data/*` (see the `ls` call above, it should match even if you download from huggingface).
## 2.1 Using the provided viewer
Basic usage:
```
python scripts/dronescapes_viewer.py data/test_set_annotated_only/ # or any of the 8 directories in data/
```
<details>
<summary> Expected output </summary>
```
[MultiTaskDataset]
- Path: '/scratch/sdc/datasets/dronescapes/data/test_set_annotated_only'
- Only full data: False
- Representations (8): [NpzRepresentation(depth_dpt), NpzRepresentation(depth_sfm_manual202204), NpzRepresentation(edges_dexined), NpzRepresentation(normals_sfm_manual202204), NpzRepresentation(opticalflow_rife), NpzRepresentation(rgb), NpzRepresentation(semantic_mask2former_swin_mapillary_converted), NpzRepresentation(semantic_segprop8)]
- Length: 116
== Shapes ==
{'depth_dpt': torch.Size([540, 960]),
'depth_sfm_manual202204': torch.Size([540, 960]),
'edges_dexined': torch.Size([540, 960]),
'normals_sfm_manual202204': torch.Size([540, 960, 3]),
'opticalflow_rife': torch.Size([540, 960, 2]),
'rgb': torch.Size([540, 960, 3]),
'semantic_mask2former_swin_mapillary_converted': torch.Size([540, 960]),
'semantic_segprop8': torch.Size([540, 960])}
== Random loaded item ==
/export/home/proiecte/aux/mihai_cristian.pirvu/.conda/envs/ngc/lib/python3.10/site-packages/numpy/core/_methods.py:215: RuntimeWarning: overflow encountered in reduce
arrmean = umr_sum(arr, axis, dtype, keepdims=True, where=where)
{'depth_dpt': tensor[540, 960] x∈[0.031, 1.000] μ=0.060 σ=0.038,
'depth_sfm_manual202204': tensor[540, 960] f16 x∈[0., 1.195e+03] μ=360.250 σ=inf,
'edges_dexined': tensor[540, 960] x∈[0.131, 1.000] μ=0.848 σ=0.188,
'normals_sfm_manual202204': tensor[540, 960, 3] f16 x∈[0.000, 1.000] μ=0.525 σ=inf,
'opticalflow_rife': tensor[540, 960, 2] f16 x∈[-0.000, 0.007] μ=0.002 σ=0.002,
'rgb': tensor[540, 960, 3] u8 x∈[0, 255] μ=68.154 σ=33.902,
'semantic_mask2former_swin_mapillary_converted': tensor[540, 960] u8 x∈[0, 7] μ=3.591 σ=3.058,
'semantic_segprop8': tensor[540, 960] u8 x∈[0, 6] μ=1.057 σ=0.916}
== Random loaded batch ==
{'depth_dpt': torch.Size([5, 540, 960]),
'depth_sfm_manual202204': torch.Size([5, 540, 960]),
'edges_dexined': torch.Size([5, 540, 960]),
'normals_sfm_manual202204': torch.Size([5, 540, 960, 3]),
'opticalflow_rife': torch.Size([5, 540, 960, 2]),
'rgb': torch.Size([5, 540, 960, 3]),
'semantic_mask2former_swin_mapillary_converted': torch.Size([5, 540, 960]),
'semantic_segprop8': torch.Size([5, 540, 960])}
== Random loaded batch using torch DataLoader ==
{'depth_dpt': torch.Size([5, 540, 960]),
'depth_sfm_manual202204': torch.Size([5, 540, 960]),
'edges_dexined': torch.Size([5, 540, 960]),
'normals_sfm_manual202204': torch.Size([5, 540, 960, 3]),
'opticalflow_rife': torch.Size([5, 540, 960, 2]),
'rgb': torch.Size([5, 540, 960, 3]),
'semantic_mask2former_swin_mapillary_converted': torch.Size([5, 540, 960]),
'semantic_segprop8': torch.Size([5, 540, 960])}
```
</details>
## 3. Evaluation for semantic segmentation
We evaluate in the paper on the 3 test scenes (unsees at train) as well as the semi-supervised scenes (seen, but
different split) against the human annotated frames. The general evaluation script is in
`scripts/evaluate_semantic_segmentation.py`.
General usage is:
```
python scripts/evaluate_semantic_segmentation.py y_dir gt_dir -o results.csv --classes C1 C2 .. Cn
[--class_weights W1 W2 ... Wn] [--scenes s1 s2 ... sm]
```
<details>
<summary> Script explanation </summary>
The script is a bit convoluted, so let's break it into parts:
- `y_dir` and `gt_dir` Two directories of .npz files in the same format as the dataset (y_dir/1.npz, gt_dir/55.npz etc.)
- `classes` A list of classes in the order that they appear in the predictions and gt files
- `class_weights` (optional, but used in paper) How much to weigh each class. In the paper we compute these weights as
the number of pixels in all the dataset (train/val/semisup/test) for each of the 8 classes resulting in the numbers
below.
- `scenes` if the `y_dir` and `gt_dir` contains multiple scenes that you want to evaluate separately, the script allows
you to pass the prefix of all the scenes. For example, in `data/test_set_annotated_only/semantic_segprop8/` there are
actually 3 scenes in the npz files and in the paper, we evaluate each scene independently. Even though the script
outputs one csv file with predictions for each npz file, the scenes are used for proper aggregation at scene level.
</details>
<details>
<summary> Reproducing paper results for Mask2Former </summary>
```
python scripts/evaluate_semantic_segmentation.py \
data/test_set_annotated_only/semantic_mask2former_swin_mapillary_converted/ \ # change this with your predictions dir
data/test_set_annotated_only/semantic_segprop8/ \
-o results.csv \
--classes land forest residential road little-objects water sky hill \
--class_weights 0.28172092 0.30589653 0.13341699 0.05937348 0.00474491 0.05987466 0.08660721 0.06836531 \
--scenes barsana_DJI_0500_0501_combined_sliced_2700_14700 comana_DJI_0881_full norway_210821_DJI_0015_full
```
Should output:
```
scene iou f1
barsana_DJI_0500_0501_combined_sliced_2700_14700 63.367 75.327
comana_DJI_0881_full 60.554 73.757
norway_210821_DJI_0015_full 37.998 45.928
overall avg 53.973 65.004
```
Not providing `--scenes` will make an average across all 3 scenes (not average after each metric individually):
```
iou f1
scene
all 60.456 73.261
```
</details>