Explorations in Texture Learning

This is the code repository for the work Explorations in Texture Learning in ICLR 2024, Tiny Papers track

To reproduce the results in this paper, simply run python3 main.py in your terminal. This will download the DTD dataset, fetch the pretrained ImageNet model, and run the texture learning analysis. When the program completes, the results will be contained in texture_object_top3.csv. With this csv file, you can create tables for the texture-object associations like this:

texture class	imagenet class top 1	effect size top 1	imagenet class top 2	effect size top 2	imagenet class top 3	effect size top 3
honeycombed	honeycomb	0.7627119	chain_mail	0.050847456	leaf_beetle	0.016949153
cobwebbed	spider_web	0.6615385	barn_spider	0.06153846	radio_telescope	0.03076923
waffled	waffle_iron	0.44705883	honeycomb	0.105882354	pretzel	0.07058824
knitted	dishrag	0.36956522	wool	0.26086956	cardigan	0.1521739
striped	zebra	0.35714287	tiger	0.16666667	velvet	0.11904762
spiralled	coil	0.32323232	maze	0.060606062	knot	0.04040404
bubbly	bubble	0.32038835	beer_glass	0.097087376	honeycomb	0.058252428
stratified	cliff	0.27083334	velvet	0.1875	stone_wall	0.09375
polka-dotted	bib	0.23762377	Windsor_tie	0.12871288	shower_curtain	0.10891089
dotted	bib	0.23762377	shower_curtain	0.12871288	wallet	0.10891089
wrinkled	velvet	0.21153846	quilt	0.17307693	wool	0.057692308
grid	window_screen	0.19626169	oscilloscope	0.11214953	shoji	0.07476635
woven	hamper	0.19626169	velvet	0.12149533	dishrag	0.093457945
stained	velvet	0.19298245	potpie	0.03508772	handkerchief	0.02631579
paisley	velvet	0.1923077	wool	0.115384616	shower_curtain	0.10576923
lacelike	handkerchief	0.1904762	velvet	0.104761906	stole	0.104761906
frilly	head_cabbage	0.1875	hoopskirt	0.09821428	velvet	0.08035714
perforated	strainer	0.18691589	space_heater	0.07476635	honeycomb	0.06542056
zigzagged	maze	0.17431192	envelope	0.12844037	quilt	0.110091746
meshed	window_screen	0.16216215	chainlink_fence	0.14414415	honeycomb	0.117117114
gauzy	shower_curtain	0.16071428	velvet	0.09821428	window_shade	0.071428575
crystalline	plastic_bag	0.16071428	honeycomb	0.09821428	pinwheel	0.071428575
braided	knot	0.16037735	hamper	0.11320755	dishrag	0.08490566
flecked	wool	0.15517241	velvet	0.0862069	cardigan	0.06896552
cracked	stone_wall	0.15178572	guillotine	0.0625	honeycomb	0.05357143
blotchy	velvet	0.1491228	ant	0.05263158	shower_curtain	0.02631579
banded	shower_curtain	0.14018692	web_site	0.093457945	Windsor_tie	0.07476635
matted	wool	0.13913043	komondor	0.07826087	wig	0.052173913
interlaced	maze	0.13392857	prayer_rug	0.08928572	buckle	0.08035714
chequered	wool	0.13274336	tray	0.097345136	web_site	0.07079646
veined	leaf_beetle	0.13157895	head_cabbage	0.096491225	walking_stick	0.05263158
pleated	window_shade	0.13043478	shower_curtain	0.12173913	velvet	0.11304348
lined	shower_curtain	0.11504425	wool	0.07964602	window_shade	0.07964602
crosshatched	window_screen	0.11504425	handkerchief	0.061946902	velvet	0.053097345
grooved	radiator	0.114035085	velvet	0.096491225	doormat	0.078947365
fibrous	hay	0.11304348	wool	0.052173913	pot	0.052173913
swirly	velvet	0.112068966	fire_screen	0.112068966	shower_curtain	0.094827585
scaly	honeycomb	0.112068966	velvet	0.060344826	tile_roof	0.060344826
freckled	lipstick	0.10810811	seat_belt	0.072072074	Band_Aid	0.072072074
studded	strainer	0.10526316	Windsor_tie	0.0877193	cuirass	0.07017544
marbled	velvet	0.097345136	cliff	0.053097345	spider_web	0.044247787
bumpy	custard_apple	0.09243698	jackfruit	0.05042017	abacus	0.033613447
potholed	geyser	0.09174312	volcano	0.08256881	manhole_cover	0.055045873
sprinkled	ice_cream	0.08547009	dough	0.05982906	confectionery	0.042735044
porous	French_loaf	0.078947365	velvet	0.05263158	honeycomb	0.05263158
pitted	pomegranate	0.06837607	doormat	0.051282052	velvet	0.051282052
smeared	mask	0.05982906	velvet	0.051282052	paintbrush	0.042735044

By default, the csv will contain the top 3 texture-object associations for each texture class. To change this number, run python3 main.py --maxk k with your desired value for k instead.

If running on a CPU, this may take some time (~20 minutes). If running on a GPU, the Dockerfile provided can be used to setup the environment with the proper packages installed.

To cite the paper:

@misc{hoak2024explorations,
      title={Explorations in Texture Learning}, 
      author={Blaine Hoak and Patrick McDaniel},
      year={2024},
      eprint={2403.09543},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknowledgements: This material is based upon work supported by, or in part by, the National Science Foundation under Grant No. CNS1946022 and Grant No. CNS2343611. Any opinions, findings, and conclusions or recommendations expressed in this publication are those of the author(s) and do not necessarily reflect the views of the National Science Foundation, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for government purposes notwithstanding any copyright notation hereon.