• RCDPC
• • To do list
• Train and Test
• • Train D-DPCC models:
  • Train lossless model for the compression of 2x downsampled coordinates:
  • Test:
  • Probable problems in testing:
• Results

RCDPC

To do list

1. Dynamic operators
2. Contextual Coding (to be tested)
   a. motion compression
3. Channel-wise autoregression
4. Rate Control

---

Train and Test

Train D-DPCC models:

python trainer.py --batch_size=4 --gpu=0 --lamb=10 --exp_name=I10 --dataset_dir='/tmp/dataset/8iVFB'

Train lossless model for the compression of 2x downsampled coordinates:

python trainer_lossless.py --dataset_dir='/home/zhaoxudong/dataset_npy'

In fact, the pretrained model is lossless_coder.pth. You probably needn't to retrain this model.

Test:

Estimate the bitrate with factorized entropy model, without practical and separate encoding and decoding process:

# python test_owlii.py --log_name='aaa' --gpu=1 --frame_count=32 --results_dir='results' --tmp_dir='tmp' --dataset_dir='/home/zhaoxudong/Owlii_10bit'
python test_8ivfb.py --log_name='aaa' --gpu=0 --frame_count=32 --results_dir='results' --tmp_dir='tmp' --dataset_dir='/tmp/dataset/8iVFB'
python test_8ivfb.py --log_name='aaa' --gpu=0 --frame_count=32 --results_dir='results' --tmp_dir='tmp' --dataset_dir='/public/DATA/zch/8iVFB'

With separate encoding and decoding process, which generates real bitstream,
and calculate encoding and decoding time.

python test_time.py --log_name='aaa' --gpu=0 --frame_count=32 --results_dir='results' --tmp_dir='tmp' --dataset_dir='/tmp/dataset/8iVFB'
python test_time.py --log_name='aaa' --gpu=0 --frame_count=32 --results_dir='results' --tmp_dir='tmp' --dataset_dir='/public/DATA/zch/8iVFB'

Probable problems in testing:

• If ./GPCC/tmc3: Permission denied:

chmod 777 ./GPCC/tmc3

• If ./GPCC/pc_error: Permission denied:

chmod 777 ./GPCC/pc_error

• The folder PCGCv2 need to be copied and in both the parent and current directory.

Results

Shown in the folder results_csv.

See details in the MPEG proposal: M60267 “[AI-3DGC] D-DPCC Test Results on 10 bit Owlii”, 2022/7.