ConDense backbone, weights, and evaluation code. This repo aims to replicate most of the main experiments covered in the ConDense paper.
./env.sh# Download data, replace {DATASET_NAME} with voc2012, ade20k, imagenet, or places205
./scripts/download/dataset_{DATASET_NAME}.sh
# Download weights, including ConDense, DINOv2, and corresponding head weights
./scripts/download/checkpoints.sh| Model | Task | Dataset | Eval mIoU/Acc (This Repo) | Reported mIoU/Acc |
|---|---|---|---|---|
| ConDense-g14 | Segmentation | VOC2012 | 85.388 | 85.1 |
| DinoV2-g14 | Segmentation | VOC2012 | 83.181 | 83.0 |
| ConDense-g14 | Segmentation | ADE20k | 53.450 | 53.6 |
| DinoV2-g14 | Segmentation | ADE20k | 48.989 | 49.0 |
| ConDense-g14 | Classification | ImageNet-1k | 90.130 | 89.6 |
| DinoV2-g14 | Classification | ImageNet-1k | 86.618 | 86.5 |
| ConDense-g14 | Classification | Places205 | 71.396 | 70.2 |
| DinoV2-g14 | Classification | Places205 | 69.515 | 67.5 |
We used a custom split of validation set for Places205, since the original split is not available.
# Segmentation
PYTHONPATH=. python ./scripts/eval_seg.py -c ./config/seg_voc2012_dinov2_standard.yaml
PYTHONPATH=. python ./scripts/eval_seg.py -c ./config/seg_ade20k_dinov2_standard.yaml
# Classification
PYTHONPATH=. python ./scripts/eval_cls.py -c ./config/cls_imagenet_dinov2_standard.yaml
PYTHONPATH=. python ./scripts/eval_cls.py -c ./config/cls_places205_dinov2_standard.yaml
# 3D Benchmarks
PYTHONPATH=. python ./scripts/eval_3d.pyYou can change the first several lines in yaml configs to switch between different backbones and weights.
- Add support for Places205 dataset
- 3D Env Docker / Set-Up Scripts
- 3D Backbone Impl and Weights
- 3D Backbone Evaluations
- Update README
- Depth Evaluations
- Online Query Demos