mbanani
DINOv2 Evaluation Results Different From Paper
Hello, thank you very much for this excellent work! I am trying to reproduce the paper results on Surface Normal Estimation using NAVI and DINOv2 B14, I using gradient accumulation (steps=4) on a single GPU, and the detailed information as well as the results are as follows.
python train_snorm.py backbone=dinov2_b14 +backbone.return_multilayer=True +output_dir="outputs/dinov2_b14/navi" +dataset=navi
This is my log and result:
but this is different from the corresponding result in Table 3 of the paper:
Additionally, when I set step=1 on a single GPU, I obtained results similar to those with step=4, and both differ from what is reported in the paper. Below is the results I got setting step=1 on a single GPU.
Is there something I have set up improperly? Thank you!
Hello, I also have this problem. Did you solve this? My results are:
and the config is:
optimizer: probe_lr: 0.0005 model_lr: 0.0 n_epochs: 10 warmup_epochs: 1.5 backbone: target: evals.models.dino.DINO dino_name: dinov2 model_name: vitb14 output: dense-cls layer: -1 dataset: target: evals.datasets.navi.NAVI path: /data/liyh/NAVI/navi_v1 image_mean: imagenet augment_train: true bbox_crop: true probe: target: evals.models.probes.SurfaceNormalHead uncertainty_aware: true head_type: dpt hidden_dim: 512 kernel_size: 3 system: random_seed: 8 num_gpus: 4 port: 12355 note: '' batch_size: 8 multilayer: true
