Training and Evaluation

All scripts should be run from the scripts/ directory.

Training

Train GPOcc on OccScanNet with the monocular single-frame setting:

bash scripts/train.sh <NUM_GPUS> <CONFIG>

Example — 4-GPU training with the default config:

bash scripts/train.sh 4 mono_dpt_bin16

This runs train_mono.py with:

• Config: config/<CONFIG>.py
• Output: work_dirs/<CONFIG>/

---

Evaluation

Monocular (single-frame) evaluation

Run evaluation on the OccScanNet test set:

bash train.sh <NUM_GPUS> <CONFIG> --evaluate

Example:

bash train.sh 4 mono_dpt_bin16_release --evaluate

The checkpoint is loaded from work_dirs/<CONFIG>/latest.pth by default. To load a specific checkpoint, set load_from in the config.

---

Embodied (streaming / incremental fusion) evaluation

Evaluate the model in the multi-frame streaming setting on EmbodiedOcc-ScanNet:

bash eval_embodied_for_mono.sh <CONFIG>

Example:

bash eval_embodied_for_mono.sh mono_dpt_bin16_release

# or 

bash scripts/eval_embodied_for_mono_vggt.sh mono_vggt_bin16_release