RSS
 

Posts Tagged ‘similarity’

Flickr rolls out neural-network-based Similarity Search

08 Mar

Flickr has announced a new feature called Similarity Search, a deep neural network-based discovery engine that is designed to uncover a wider range of related image results than is possible under the current search system. It lets users search for a photo and use the results to dig deeper for additional similar photos.

Powered by Yahoo’s computer vision technology, this new function expands the depth and level of search results, and not coincidentally, will derive additional advantages to users of Flickr’s 13 billion-plus photo database. To use similarity search, just hover over a photo on the search result page to reveal a three-dot (“…”) Similarity Pivot button at the upper right corner. That exposes a menu giving you the option to search for additional similar photos.

Neural networks are used to map raw pixels of a photo into relevant tags. Credit: Flickr

This spells more potential exposure for both aspiring and professional photographers, where search results can go far beyond anything that may be obvious, popular or trending. Similar to a web searching or browsing experience, users will likely encounter a wider array of different, unexpected, and hopefully valuable results to choose from without having to conduct increasingly elaborate text searches.

‘One of those major problems is that of discovery,’ said Clayton Mellina, Flickr’s software developer engineer, in the blog post explaining the new tech. ‘We understand that the value in our photo corpus is only unlocked when the community can find photos and photographers that inspire them, so we strive to enable the discovery and appreciation of new photos.’

Points nearer to the query image are semantically similar; points further away are not. Credit: Flickr

What is similarity?

While similarity can mean the same person or thing or color or style, Flickr goes into deep neural network territory with semantic similarity — similarity based on the content of photos to help users discover new images and better use Flickr’s tools to find what they’re looking for from its huge database. Flickr has some experience with neural networks and already uses them for object recognition, NSFW prediction, and aesthetic quality. In addition, neural networks power Flickr’s style filters, which allow filtering by styles like ‘minimalist’ or ‘patterns.’
 

Because visual searching often does not translate well to text, the new Similarity Pivot button now gives users a way to more deeply explore visual concepts. ‘With similarity, we are understanding the photo abstractly with specific aspects that are related to the visual experience and layout of the scene,’ Mellina said.

Similarity search rolls out today to all Flickr users, and results will be available immediately. Mellina says that the Flickr collection is still in the process of being indexed, but that indexing should be completed by the end of this week. 

Similarity search will reveal only public photos, so there should be no security issues, Mellina says.

Articles: Digital Photography Review (dpreview.com)

 
Comments Off on Flickr rolls out neural-network-based Similarity Search

Posted in Uncategorized

 

Ten important differences between Panasonic’s GH5 and GH4 (and one key similarity)

16 Jan

In-body IS

Panasonic’s new GH5 flagship will be hitting the streets soon, joining the GH4 in the company’s line of video-centric cameras. With our eye on video features, we take a look at the ten biggest differences between these two cameras.

Although this article is mainly looking at video shooting, the biggest physical change between the GH5 and its predecessor promises benefits both for stills and video shooters.

The addition of five-axis image stabilization is a major step forwards for the GH series. Using Panasonic’s latest version of its I.S. system, the camera will combine internal stabilization with in-lens stabilization when shooting with own-brand lenses. For video shooters, this includes a digital stabilization element for dealing with larger or faster translational movements than physical stabilization can correct for.

Although most serious film makers won’t consider in-body stabilization to be an effective replacement for stabilization hardware, it will significantly extend the GH5’s appeal for run-and-gun video, where a little camera shake just adds a little vérité atmosphere.

Full width 4K video

Using the full width of the sensor has three hugely significant benefits. First, it means that all the lenses designed for the Four Thirds format provide the angle of view that they were designed to give (putting to one side the slight crop as you move from the 4:3 aspect ratio down to 16:9). This makes it easier to find lenses that will give a wide-angle view of the world. This use of the full sensor width also means that focal length reducers (such as Speed Boosters) with Super 35 lenses will give the same field of view as they were designed to give.

Second, using the full width of the sensor means using a larger area of the sensor, which means more light capture and better noise performance. Using this larger region should mean up to a 1/2EV improvement over the cropped region the GH4 used for UHD 4K. Add to this improved noise reduction algorithms (that seem particularly effective for combating temporal noise, based on what we’ve seen), and the GH5 should be a significantly better performer as light falls.

Finally, there’s the resolution benefit that comes from oversampling. Capturing the sensor’s full resolution and then downsampling allows the capture of finer detail than is possible with just using a 4096 or 3840 x 2160 region of the sensor.

Internal 10-bit recording

The addition of ten bit internal video capture is a big deal. To really understand why, we’ve gone into a bit more depth in our First Impressions Review, but at its most basic, the move to 10-bit recording means 4x the number of ‘steps’ available in each channel: 64x as much data in total. This means much more subtle gradation can be captured, meaning more tolerance for subsequent grading and processing. On top of this, the GH5 can capture with 4:2:2 chroma sub-sampling, effectively meaning twice the color resolution of the GH4.

The GH4 was one of the few cameras we’ve encountered to allow 10-bit output to an external recorder, so it’s a big deal that 10-bit 4:2:2 footage can be recorded internally on the GH5 (though you’ll still need an external recorder if you want to use the camera’s highest frame rates).

The increased grading flexibility should be especially useful for any users who buy the optional V-Log upgrade, since it gives more bit-depth to compress the camera’s dynamic range into: decreasing the risk of banding and posterization when the footage is pulled back into a end-use color space.

60p 4K capture

The GH5 is only the second camera we’ve yet encountered that can shoot 4K footage at above 30 frames per second. The GH5 gains the ability to shoot UHD 4K video at 48, 50 and 59.94p.

60p (and 50p if you’re shooting for PAL output) is great for capturing fast movement, rather than giving the blurred impression of movement that 24p brings. Several shooters in the office have become big fans of the look. Alternatively, of course, It can be slowed back down to 24p, giving UHD footage at 1/2.5th speed. Either way, it’s another creative option in your toolbox to use (or not) as suits your style.

Pro video tools

Dig a little deeper into the GH5 and its videographer-focused features and you’ll find it moves beyond the already impressive array offered by the GH4. In addition to existing features such as time code, SMPTE bars and the option to limit output to broadcast-safe data values, the GH5 add a series of features and capabilities.

For instance, the GH5 can present Waveform and Vectorscope representations of the scene being shot. This ability to visualize your shot using video-industry standard tools is a welcome step forward, and the first time we’ve seen them in a hybrid stills/video camera.

And, while the GH4 could already be upgraded to shoot with a Log profile (another first for a camera coming through our office), the GH5 VLog upgrade adds the ability to load Look Up Tables (LUTs) onto the camera, so that the camera can show a ‘faux-graded’ preview while shooting in the harder-to-interpret Log gamma and color mode.

The other video-friendly feature that stood out to us was the ability to pre-define a series of up to three focus distances before you start shooting, meaning that the camera can rack focus between these points in a controlled manner. It means it’s possible to refocus smoothly without a dedicated focus puller or any of the uncertainty of autofocus.

User interface (Menus and My Menu)

The GH5’s user interface has received a significant overhaul. Enough of it remains the same that existing users will be able to start shooting straight away but the menus have been revamped to make them easier to navigate and remember.

In addition to the better menu categorization and improved location indication, Panasonic has also added a ‘My Menu’ tab, meaning that users can get fast access their most-used settings. These settings can then be saved onto an SD card, meaning this customization doesn’t need to be repeated every time you switch cameras.

On top of these menu changes, Panasonic has made a host of small tweaks that make a big difference to everyday operation. One such example is that the camera now remembers different focus settings for its stills and video modes, meaning that you don’t have to constantly reconfigure the camera to the differing demands of still and video work.

Secure, full-size HDMI socket

Another seemingly small change to the camera’s body is the move to using a full sized HDMI socket, acknowledging the likelihood that the camera will be connected to an external monitor or recorder.

On top of this, Panasonic has included a cable retainer that screws into the camera body, in the box. This greatly decreases not only the risk of a cable coming loose, mid shoot but also significantly decreases the risk of breaking the HDMI socket, which is not unheard-of on hard-working video cameras.

XLR accessory

One challenge when shooting video on still cameras is how to record high quality sound. Most still cameras include a simple audio-in mini jack, meaning that if you want to record using proper XLR-connected microphones you need to use an third party adapter or record sound off camera and sync in post.

On the GH4, Panasonic solved this with its optional YAGH unit, which included XLR inputs as well as four SDI outputs. A good solution to be sure, but it cost $ 2000!

On the GH5 Panasonic is offering what appears to be a reasonable compromise: the DMW-XLR1 microphone accessory. The XLR1 mounts to the camera’s hot shoe, includes two XLR inputs, physical switches and dials similar to what you would find on a pro video camera, and records 96kHz/24-bit sound. It even has a dummy shoe mount on top in case you need to mount another accessory, such as a wireless receiver, above. It doesn’t include SDI connectors, but we’re guessing that most GH5 users will be satisfied with HDMI-out.

We don’t have pricing information yet, but we’d like to think it will cost less than $ 2000.

Auto ISO in M movie shooting

It’s a little thing, but one we very much welcome: with the GH5, Panasonic has added the ability to shoot with Auto ISO in Manual video mode.

Allowing the camera control settings in a ‘manual’ mode may seem like a head-scratcher at first but, particularly for run-and-gun shooting, it can be a really useful combination. It means that the camera operator can set their aperture and shutter speed to maintain the depth-of-field and sense of movement that they want to convey, while trusting the camera to maintain the correct exposure even if the lighting changes.

Offering exposure compensation means that the user retains control over what ‘correct’ exposure is.

Viewfinder

The GH5’s rear screen has been updated but it’s the viewfinder that’s most immediately striking. Its step up to 3.68m dots means it is able to offer 1280 x 960 pixel resolution: a 25% increase in linear resolution compared with the GH4. This may not sound like a major benefit for video shooters but a to-the-eye shooting stance gives a much steadier grip for hand-held shooting than holding the camera at waist level.

The excellent, detailed viewfinder, along with built-in stabilization is only likely to encourage shooting this way, even if only for a shot here and there.

The rear screen moves from 1.04m dots (720 x 480 pixels) to 1.62m dots (900 x 600 pixels). The GH5’s panel should also be more efficient or brighter as it uses an RGB pattern with white pixels interspersed into the array, rather than the conventional RGB pattern used on the GH4.

Battery and media

For all the things that have changed, Panasonic has done a good job keeping as much as possible the same. [REC] button aside, the new camera’s ergonomics are broadly unchanged, compared with its predecessor. But it’s the decision to use the same batteries and media type that’s likely to please existing GH4 users.

The GH5’s promised 400Mbps capture (due in a summer 2017 firmware update) will be too much for most current SD cards, which could have prompted a move to a different capture media. Instead, Panasonic has stayed true to the SD format and added a second slot, compatible with the forthcoming V60 class cards that will guarantee to support the appropriate data rates.

Sticking to the same DMW-BLF19 battery as the GH3 and 4 means any money invested in spare batteries won’t be wasted. The new camera is rated at 410 shots per charge, down from 530 on the GH4 but it’s unlikely many users would have welcomed the larger body and need for a whole set of extra batteries that a change in power supply would have necessitated.

Articles: Digital Photography Review (dpreview.com)

 
Comments Off on Ten important differences between Panasonic’s GH5 and GH4 (and one key similarity)

Posted in Uncategorized