Picking bug

Version: 4.0.0-rc1

Steps to reproduce

• Create a ScatterplotLayer
• Zoom and pan

Observed result

Once per minute or so, it crashes with the following error:

Uncaught TypeError: Cannot read property 'model' of null
    at ScatterplotLayer.setUniforms (webpack:///./~/deck.gl/dist/lib/layer.js?:767:21)
    at ScatterplotLayer.getPickingInfo (webpack:///./~/deck.gl/dist/lib/layer.js?:201:14)
    at ScatterplotLayer.pickLayer (webpack:///./~/deck.gl/dist/lib/layer.js?:590:19)
    at eval (webpack:///./~/deck.gl/dist/lib/draw-and-pick.js?:152:20)
    at Array.forEach (native)
    at eval (webpack:///./~/deck.gl/dist/lib/draw-and-pick.js?:141:20)
    at glContextWithState (webpack:///./~/luma.gl/dist-es6/webgl/context.js?:188:13)
    at pickLayers (webpack:///./~/deck.gl/dist/lib/draw-and-pick.js?:63:32)
    at LayerManager.pickLayer (webpack:///./~/deck.gl/dist/lib/layer-manager.js?:188:42)
    at DeckGL._onMouseMove (webpack:///./~/deck.gl/dist/react/deckgl.js?:215:45)
    at EventsProxy.mousemove (webpack:///./~/luma.gl/dist-es6/core/event.js?:352:24)
    at HTMLCanvasElement.eval (webpack:///./~/luma.gl/dist-es6/core/event.js?:160:26)

I couldn’t find a reliable way to reproduce this, but it happens frequently.

[ibgreen]

@dcposch - Yes this is a very interesting idea, we have actually considered this. As an example we have played with the idea of using separate invisible support layers (with the same data) to provide wider hit areas for picking (e.g. we have a z coordinate based Voronoi layer that we have experimented with for point picking - it works exactly as you describe by rendering a “cone” facing away from the viewer for each point, so the z buffer automatically determines which data point is nearest any screen pixel).

Improving picking could be a good theme for deck.gl v4.1 / v5. It needs some thought though - for really small objects it makes sense to have a wider target area but for bigger objects I think being pixel precise is desirable, at least in some apps.

In some cases there might be other good solutions that can provide good results (e.g. for the icon layer one could consider supporting a second bit mask used during picking to avoid not being able to pick inside transparent parts of icons?).

[dcposch]

@Pessimistress yeah, that’s a cool idea. dunno if you have to “traverse in a spiral”, but regardless of the exact way it’s done, once you’ve already rendered the whole picking buffer, it should be negligible extra computation to:

• check a neighborhood around a given pixel coordinate
• see if there are any objects within some small radius
• if so, return the closest object