Experiments on ShapeNet

Data

We use ShapeNet data and their renderings, as provided by R2N2.

Run

datasets/shapenet/download_shapenet.sh

to download R2N2, and the train/val/test splits. You also need the original ShapeNet Core v1 & binvox dataset, which require registration before downloading.

Preprocessing

python tools/preprocess_shapenet.py \
--shapenet_dir /path/to/ShapeNetCore.v1 \
--shapenet_binvox_dir /path/to/ShapeNetCore.v1.binvox \
--output_dir ./datasets/shapenet/ShapeNetV1processed \
--zip_output

The above command preprocesses the ShapeNet dataset to reduce the data loading time. The preprocessed data will be saved in ./datasets/shapenet and will be zipped. The zipped output is useful when training in clusters.

Training

python tools/train_net_shapenet.py --num-gpus 8 \
--config-file configs/shapenet/voxmesh_R50.yaml

When --copy_data, the preprocessed zipped data from above will be copied to a local /tmp directory. This is particularly useful when training on remote clusters, as it reduces the io time during training

Testing and Evaluation

python tools/train_net_shapenet.py --eval-only --num-gpus 1 \
--config-file configs/shapenet/voxmesh_R50.yaml \
MODEL.CHECKPOINT shapenet://voxmesh_R50.pth

The output of the evaluation produces the results as shown in Table 2 of our paper. To evaluate under the Pixel2Mesh protocol, as in Table 1 of our paper, add --eval-p2m.

Models

Mesh R-CNN	Pixel2Mesh	SphereInit
voxmesh_R50.pth	pixel2mesh_R50.pth	sphereinit_R50.pth

Note that we release only the light and pretty for both our and the baseline models.

Performance

Scale-normalized Protocol

Performance of our model on ShapeNet test set under the scale-normalized evaluation protocol (as in Table 2 of our paper).

category	#instances	chamfer	normal	F1(0.1)	F1(0.3)	F1(0.5)
bench	8712	0.120899	0.657536	42.4005	86.0036	95.128
chair	32520	0.183693	0.712362	31.6906	79.8275	92.0139
lamp	11122	0.413965	0.672992	30.5048	70.3449	84.5068
speaker	7752	0.253796	0.730829	24.8335	74.6606	88.237
firearm	11386	0.168323	0.621439	47.2251	85.271	93.8171
table	40796	0.148357	0.75642	42.249	86.2039	94.1623
watercraft	9298	0.224168	0.642812	30.0589	75.5332	89.9764
plane	19416	0.187465	0.684285	39.009	80.998	92.1069
cabinet	7541	0.111294	0.75122	34.8227	86.9346	95.371
car	35981	0.107605	0.647857	29.6397	85.7925	96.2938
monitor	5256	0.218032	0.779365	27.2531	77.2979	90.904
couch	15226	0.144279	0.72302	27.5734	81.684	94.3294
cellphone	5045	0.121504	0.850437	42.9168	88.9888	96.1367
total	210051	0.184875	0.710044	34.629	81.5031	92.5372
per-instance	210051	0.171189	0.70275	34.9372	82.4107	93.1323

Pixel2Mesh Protocol

Performance of our model on ShapeNet test set under the pixel2mesh evaluation protocol (as in Table 1 of our paper). To evaluate under this protocol, add the --eval-p2m flag.

category	#instances	chamfer	normal	F1(0.0001)	F1(0.0002)
bench	8712	0.000295252	0.657508	73.4681	84.4999
chair	32520	0.000400415	0.712348	66.5227	79.3634
lamp	11122	0.000788915	0.673057	60.1057	71.7711
speaker	7752	0.000582152	0.730797	59.9974	73.8792
firearm	11386	0.000357016	0.621438	75.9761	85.5111
table	40796	0.000342991	0.756442	76.0776	85.4878
watercraft	9298	0.000449061	0.642791	62.808	76.5464
plane	19416	0.000313141	0.684333	75.8104	85.3897
cabinet	7541	0.000293613	0.751306	72.6302	84.7327
car	35981	0.000240585	0.647896	71.8118	85.5155
monitor	5256	0.000470965	0.779397	64.2917	77.8422
couch	15226	0.000355369	0.723013	64.1388	79.327
cellphone	5045	0.000280397	0.850456	77.3011	87.8698
total	210051	0.000397682	0.71006	69.303	81.3643
per-instance	210051	0.000368317	0.702768	70.4479	82.3373