Awesome

Minimal LSeg impl in PyTorch

This is a minimal version of the dense (per-pixel) CLIP feature extractor proposed in Language-driven semantic segmentation at ICLR 2022.

Note: I am not an author on the paper. I merely took their original implementation and extracted the core components therein.

Install requirements

Install pytorch

First, install a recent pytorch release by following instructions on https://pytorch.org/

Install OpenAI CLIP

Now, install the OpenAI CLIP package by running

pip install git+https://github.com/openai/CLIP.git

Build package

Build the lseg package by running

python setup.py build develop

Try it out (examples)

Download pretrained weights

Download the weights of the pretrained model from this OneDrive link.

Copy this over to the examples/checkpoints directory.

Note: Source images used in the examples are from Wikimidea and the original LSeg repo.

Example 1: extract_lseg_features.py

cd examples
python extract_lseg_features.py

Commandline options supported

usage: extract_lseg_features.py [-h] [--checkpoint-path STR|PATH]
                                [--backbone STR] [--num-features INT]
                                [--arch-option INT] [--block-depth INT]
                                [--activation STR] [--crop-size INT]
                                [--query-image STR|PATH] [--prompt STR]
╭─ arguments ────────────────────────────────────────────────────────────────╮
│ -h, --help              show this help message and exit                    │
│ --checkpoint-path STR|PATH                                                 │
│                         (default:                                          │
│                         lseg-minimal/examples/checkpoi…                    │
│ --backbone STR          (default: clip_vitl16_384)                         │
│ --num-features INT      (default: 256)                                     │
│ --arch-option INT       (default: 0)                                       │
│ --block-depth INT       (default: 0)                                       │
│ --activation STR        (default: lrelu)                                   │
│ --crop-size INT         (default: 480)                                     │
│ --query-image STR|PATH  (default:                                          │
│                         lseg-minimal/images/teddybear.…                    │
│ --prompt STR            (default: teddy)                                   │
╰────────────────────────────────────────────────────────────────────────────╯

This script should produce an output that looks like:

Example 2: lseg_openset_seg.py

This example loads an image, takes an arbitrary set of class labels specified as comma separated text WITHOUT ANY WHITE SPACES, and performs open-set segmentation.

cd examples
python lseg_openset_seg.py

Commandline options supported

usage: lseg_openset_seg.py [-h] [--checkpoint-path STR|PATH] [--backbone STR]
                           [--num-features INT] [--arch-option INT]
                           [--block-depth INT] [--activation STR]
                           [--crop-size INT] [--query-image STR|PATH]
                           [--segclasses STR]

╭─ arguments ────────────────────────────────────────────────────────────────╮
│ -h, --help              show this help message and exit                    │
│ --checkpoint-path STR|PATH                                                 │
│                         (default:                                          │
│                         lseg-minimal/examples/checkpoi…                    │
│ --backbone STR          (default: clip_vitl16_384)                         │
│ --num-features INT      (default: 256)                                     │
│ --arch-option INT       (default: 0)                                       │
│ --block-depth INT       (default: 0)                                       │
│ --activation STR        (default: lrelu)                                   │
│ --crop-size INT         (default: 480)                                     │
│ --query-image STR|PATH  (default:                                          │
│                         lseg-minimal/images/cat.jpg)                       │
│ --segclasses STR        (default: plant,grass,cat,stone)                   │
╰────────────────────────────────────────────────────────────────────────────╯

When run with the default options, this should produce the following image