Skip to content

Latest commit

 

History

History
112 lines (75 loc) · 4.33 KB

INFERENCE.md

File metadata and controls

112 lines (75 loc) · 4.33 KB
Raw

Inference

What is palette? Why is the output a "colored image"? How do I make those input masks that look like color images? See PALETTE.md.

  1. Set up the datasets following GETTING_STARTED.md.
  2. Download the pretrained models either using ./scripts/download_models.sh, or manually and put them in ./saves (create the folder if it doesn't exist). You can download them from [GitHub] or [Google Drive].

All command-line inference are accessed with eval.py. See RESULTS.md for an explanation of FPS and the differences between different models.

Usage

python eval.py --model [path to model file] --output [where to save the output] --dataset [which dataset to evaluate on] --split [val for validation or test for test-dev]

See the code for a complete list of available command-line arguments.

Examples: (--model defaults to ./saves/XMem.pth)

DAVIS 2017 validation:

python eval.py --output ../output/d17 --dataset D17

DAVIS 2016 validation:

python eval.py --output ../output/d16 --dataset D16

DAVIS 2017 test-dev:

python eval.py --output ../output/d17-td --dataset D17 --split test

YouTubeVOS 2018 validation:

python eval.py --output ../output/y18 --dataset Y18

Long-Time Video (3X) (note that mem_every, aka r, is set differently):

python eval.py --output ../output/lv3 --dataset LV3 --mem_every 10

Getting quantitative results

We do not provide any tools for getting quantitative results here. We used the followings to get the results reported in the paper:

(For the Long-Time Video dataset, point --davis_path to either long_video_davis or long_video_davis_x3)

You can also use my own script which evaluates much faster and produces identical results: vos-benchmark.

On custom data

Structure your custom data like this:

├── custom_data_root
│   ├── JPEGImages
│   │   ├── video1
│   │   │   ├── 00001.jpg
│   │   │   ├── 00002.jpg
│   │   │   ├── ...
│   │   └── ...
│   ├── Annotations
│   │   ├── video1
│   │   │   ├── 00001.png
│   │   │   ├── ...
│   │   └── ...

We use sort to determine frame order. The annotations do not have have to be complete (e.g., first-frame only is fine). We use PIL to read the annotations and np.unique to determine objects. PNG palette will be used automatically if exists.

Then, point --generic_path to custom_data_root and specify --dataset as G (for generic).

Multi-scale evaluation

Multi-scale evaluation is done in two steps. We first compute and save the object probabilities maps for different settings independently on hard-disks as hkl (hickle) files. Then, these maps are merged together with merge_multi_score.py.

Example for DAVIS 2017 validation MS:

Step 1 (can be done in parallel with multiple GPUs):

python eval.py --output ../output/d17_ms/720p --mem_every 3 --dataset D17 --save_scores --size 720
python eval.py --output ../output/d17_ms/720p_flip --mem_every 3 --dataset D17 --save_scores --size 720 --flip

Step 2:

python merge_multi_scale.py --dataset D --list ../output/d17_ms/720p ../output/d17_ms/720p_flip --output ../output/d17_ms_merged

Instead of --list, you can also use --pattern to specify a glob pattern. It also depends on your shell (e.g., zsh or bash).

Advanced usage

To develop your own evaluation interface, see ./inference/ -- most importantly, inference_core.py.