HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts (ICLR 2025)

HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts
By Hongjun Wang, Sagar Vaze, and Kai Han.

Prerequisite 🛠️

First, you need to clone the HiLo repository from GitHub. Open your terminal and run the following command:

git clone https://github.com/Visual-AI/HiLo.git
cd HiLo

We recommend setting up a conda environment for the project:

conda create --name=hilo python=3.9
conda activate hilo
pip install -r requirements.txt

Running 🏃

Config

Set paths to datasets and desired log directories in config.py

Datasets

We use DomainNet and our created Semantic Shift Benchmark Corruption (SSB-C) datasets:

• DomainNet
• SSB-C (Personal OneDrive / HKU Data Repository)

Checkpoints

Download the checkpints of HiLo for different datasets / combination (only used during evaluation).

• Google Drive

Scripts

Eval the model

python -m methods.ours.evaluate \
    --dataset_name domainnet \
    --src_env 'real' \
    --aux_env 'painting' \
    --tgt_env 'sketch' \
    --checkpoint_path /path/to/checkpoint.pt \
    --task_type 'A_L+A_U+B->A_U+B+C'

To reproduce all main results in the paper, just change the name (dataset_name), (aux_env), (tgt_env) and its corresponding path (checkpoint_path) to the pretrained model you downloaded from the above link.

Train the model:

bash scripts/mi_pmtrans/domainnet.sh 0
bash scripts/mi_pmtrans/ssbc.sh 0

Just be aware to make necessary changes (e.g., PYTHON, SAVE_DIR, WEIGHTS_PATH, etc).

Results

DomainNet results:

Real+Painting

Methods	Real (All)	Real (Old)	Real (New)	Painting (All)	Painting (Old)	Painting (New)
RankStats+	34.1	62.0	19.7	29.7	49.7	9.6
UNO+	44.2	72.2	29.7	30.1	45.1	17.2
ORCA	31.9	49.8	23.5	28.7	38.5	7.1
GCD	47.3	53.6	44.1	32.9	41.8	23.0
SimGCD	61.3	77.8	52.9	34.5	35.6	33.5
HiLo (Ours)	64.4	77.6	57.5	42.1	42.9	41.3

Real+Sketch

Methods	Real (All)	Real (Old)	Real (New)	Sketch (All)	Sketch (Old)	Sketch (New)
RankStats+	34.2	62.0	19.8	17.1	31.1	6.8
UNO+	43.7	72.5	28.9	12.5	17.0	9.2
ORCA	32.5	50.0	23.9	11.4	14.5	7.2
GCD	48.0	53.8	45.3	16.6	22.4	11.1
SimGCD	62.4	77.6	54.6	16.4	20.2	13.6
HiLo (Ours)	63.3	77.9	55.9	19.4	22.4	17.1

Real+Quickdraw

Methods	Real (All)	Real (Old)	Real (New)	Quickdraw (All)	Quickdraw (Old)	Quickdraw (New)
RankStats+	34.1	62.5	19.5	4.1	4.4	3.9
UNO+	31.1	60.0	16.1	6.3	5.8	6.8
ORCA	19.2	39.1	15.3	3.4	3.5	3.2
GCD	37.6	41.0	35.2	5.7	4.2	6.9
SimGCD	47.4	64.5	37.4	6.6	5.8	7.5
HiLo (Ours)	58.6	76.4	52.5	7.4	6.9	8.0

Real+Clipart

Methods	Real (All)	Real (Old)	Real (New)	Clipart (All)	Clipart (Old)	Clipart (New)
RankStats+	34.0	62.4	19.4	24.1	45.1	6.2
UNO+	44.5	66.1	33.3	21.9	35.6	10.1
ORCA	32.0	49.7	23.9	19.1	31.8	4.3
GCD	47.7	53.8	44.3	22.4	34.4	16.0
SimGCD	61.6	77.2	53.6	23.9	31.5	17.3
HiLo (Ours)	63.8	77.6	56.6	27.7	34.6	21.7

Real+Infograph

Methods	Real (All)	Real (Old)	Real (New)	Infograph (All)	Infograph (Old)	Infograph (New)
RankStats+	34.2	62.4	19.6	12.5	21.9	6.3
UNO+	42.8	69.4	29.0	10.9	15.2	8.0
ORCA	29.1	47.7	20.1	8.6	13.7	7.1
GCD	41.9	46.1	39.0	10.9	17.1	8.8
SimGCD	52.7	67.0	44.8	11.6	15.4	9.1
HiLo (Ours)	64.2	78.1	57.0	13.7	16.4	11.9

SSB-C results:

CUB-C

Methods	Original (All)	Original (Old)	Original (New)	Corrupted (All)	Corrupted (Old)	Corrupted (New)
RankStats+	19.3	22.0	15.4	13.6	23.9	4.5
UNO+	25.9	40.1	21.3	21.5	33.4	8.6
ORCA	18.2	22.8	14.5	21.5	23.1	18.9
GCD	26.6	27.5	25.7	25.1	28.7	22.0
SimGCD	31.9	33.9	29.0	28.8	31.6	25.0
UniOT	27.5	29.3	26.8	27.3	33.2	22.5
HiLo (Ours)	56.8	54.0	60.3	52.0	53.6	50.5

Scars-C

Methods	Original (All)	Original (Old)	Original (New)	Corrupted (All)	Corrupted (Old)	Corrupted (New)
RankStats+	14.8	20.8	7.8	11.5	22.6	1.0
UNO+	22.0	41.8	7.0	16.9	29.8	4.5
ORCA	19.1	28.7	11.2	15.0	22.4	8.3
GCD	22.1	35.2	20.5	21.6	29.2	10.5
SimGCD	26.7	39.6	25.6	22.1	30.5	14.1
UniOT	24.3	37.5	22.3	22.9	31.4	13.7
HiLo (Ours)	39.5	44.8	37.0	35.6	42.9	28.4

FGVC-C

Methods	Original (All)	Original (Old)	Original (New)	Corrupted (All)	Corrupted (Old)	Corrupted (New)
RankStats+	14.4	16.4	14.5	8.3	15.6	5.0
UNO+	22.0	33.4	15.8	16.5	25.2	8.8
ORCA	17.6	19.3	16.1	13.9	17.3	10.1
GCD	25.2	28.7	23.0	21.0	23.1	17.3
SimGCD	26.1	28.9	25.1	22.3	23.2	21.4
UniOT	27.3	29.8	22.5	21.6	23.5	19.6
HiLo (Ours)	44.2	50.6	47.4	31.2	29.0	33.4

Citing this work

If you find this repo useful for your research, please consider citing our paper:

@inproceedings{wang2025hilo,
  title={HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts},
  author={Wang, Hongjun and Vaze, Sagar and Han, Kai},
  booktitle={ICLR},
  year={2025}
}