Release dataset version 0.0.4

[nottombrown]

This PR does the following:

• Corrects dataset issues caused by faulty tasker pipeline (described in #59 )
• Allows researchers to inspect the raw labels given by the taskers.
• Filters train and test dataset to exclude images that are truncated. Note that this reduces the size of the IID train and test sets to 125 images each. However, the excluded images can still be used for training in the large extras dataset

Researchers can upgrade to the new dataset by pulling the repo and running the following command

pip install -e bird-or-bicycle --upgrade

Bad bird datapoints that are now removed

-3b2f2608d3f04aa1
-79ac2f60e526e683
-1ab16de89590e4c5
-99636d7b4ea3e233
-93cd5a5b526d5e15
-ff721aa6906a239b
-7aba5c4ae8c57916
-d84af53461a42947
-85785cc1a59e8412
-3aa31e2c8960ec33
-7485d8abf0b9297b
-954c63e11aa1232e
-ef14662cd1c2d115
-f616e4fc0d3836b3
-2e8cd9f60546ac55
-001c9606f49c59c4
-598820468ca2957d
-d0a53873d9b5045c
-6b3797dec5bea3c7
-cae73b8b09db3e93
-e9e1abec98923fbe
-386aca1b6dbc7ab3
-a69a9641ac3c1fad
-397906f964c1ba37
-466724fc1a2f5ea8
-b0cf7e6ba0f7c941
-4fba9527aa917ef8
-2f493ee0e8482bd1
-17fbc19521a99b25
-8366a15ab29977f8
-41a87bc25d7eafd0
-defb4c991b77aa01
-21f2aff1121d13be
-e0726c3ba7829ff4
-4b347ef789b71f2f
-d1125249cd57e7a7
-8e385d43da95930c
-68775c9d8051dcce
-bcab0463d981755e
-8a084a78ead09bbb
-9ae97d6e4190b7b4
-9c59fb41857ecfda
-603c2f061ce7dbd3
-1b0fb5b533b24537
-1ba42f8e412b76b3
-72460159badf7b64

Bad bicycle datapoints that are now removed

-8ae043c7b7448219
-80a1a3e948880d93
-189467f4fea0a56a
-c1338874b4a98446
-842121bda37f25a2
-d6bcc9441f2e2317
-63390b707dbb7022
-91370c5768f8dcfa
-410b9c96ebc154ef
-a585f18d33cb9ab5
-1c7ee7e854fa4d19
-8a650416f08b3860
-2886fdcc63085b01
-9dc74e41a54c6688
-33f3dd75e994bc9b
-ed55ba03dfdd99d7
-5fd811193609c657
-ef5ad509b85abb72
-37e6a15259882305
-ede99f867f9f6f2b
-b879da8e1ade53a1
-315034ca9726da08
-717815eaf2ef2ca4
-394bd46368ac2e40
-5aea87bab366c4ad
-8e7b9ef0c246a18b
-055b2035b7557d4f
-802846b84af1e218
-7c9064473e6a6a78
-f44d753798ef7bf5
-cf74a4afc09463de
-626281361aa05d80
-3de4c2b947bcb15d
-e2c737ad4e6c742b
-05aea1cd2c7641b3
-cac51471e70ed382
-87241e2568cd46aa
-8c20158094acab48
-9dca3efa0d64899e
-977074739a92ea75
-aafaf3228318cc7b
-f451d2ea5f218cee
-3830b17dc5980dd8
-7a4f3c534956200f
-8354f24d0006eacf
-79b6814d3471cdd7
-d9084e9ba4e51173
-079d91186621a4d9
-96f73c14700f8944

[carlini]

How did you find these images?

[nottombrown]

@carlini - These bad images were due to the issue with the tasker pipeline. I filed this as an incident and will directly export data from the tasker database myself in the future.

We had been aware of the issue earlier, but I had failed to commit my changes that used the updated csv file.

[nottombrown]

I have manually verified that all the images in the following dataset folders appear unambiguous to me and are not truncated:

• ✅ bird train
• ✅ bicycle train
• ✅ bird test
• ✅ bicycle test

[nottombrown]

@carlini can I get a quick review?

[carlini]

Let me take a look

[carlini]

It all seems fine