We were asked to find a solution to creating crowds that would allow the director complete flexibility to direct behavior and the freedom of a dynamic camera, but ultimately produce a crowd that would seamlessly extend the existing crowds.

There were a number of ways to approach the project.

The most obvious is with a locked off camera with real people in shot. In order to do this we would have to spend a huge amount of time moving crowds of people around to different areas of the shot and change their wardrobes and performances to match the area of the plate they were to perform in. The amount of time required to do this would have been extremely prohibitive.

In addition, Ian wanted a number of helicopter shots that would show off the scale of the crowds. The look needed to feel like a herd of zebra or Wildebeese running and changing directions.

BlackGinger had been doing a lot of research into crowd systems for about a year prior to this production. Brian Goodwin our VFX Technical Director had completed a production for Greg Grey of Velocity, where they had created digital crowds for a Total Rugby spot using a system developed within Houdini, prior to joining us. After extensive research into the best crowd systems on the market we decided to invest in Massive. It was production proven on films such as Lord Of The Rings, 300, Happy Feet, World Trade Center and I Robot to name a few. Also, in the commercials world Frank Budgen had used the technology on his Cannes Grand Prix commercial for Sony Playstation Mountain and Plaza Films used it on Carlton Draughts Big Ad. Once we had acquired the software, we spent a huge amount of time and resources in training and pipeline development, as this software was not your average, buy it for the job and do it.

To implement this tool and reshape our production pipeline we invested in a bunch of new skills, setting up a complete programming and R&D department. We needed to develop a series of our own applications that would allows us to transport information such as camera data, agent models, rigs, animation etc between all our tools seamlessly. Wed develop shot elements in XSI, or take camera and terrain information from our 3d tracking software, PFTrack, into Houdini, Massive and Flame directly. This would allow the Digital Actors to realistically react to the ground and allow us to clean up the shots or extend them more easily.

Once all of this was in place we felt we were able to now take on a production and deliver it to a photoreal standard.

It was shortly after this that Sanlam herds came into the studio. We did a number of tests for Ian that showed proof of concept to him and the agency. When they could not tell the difference between the CG and the live action, the job was on.

The process.

In order to do a production of this nature we needed to get the best possible data.

PreViz.

Once the location for the commercial was chosen, we went and took survey photographs with the director and discussed extensively how he saw each shot. We then created a low-resolution edit with all the type of camera moves that Ian wanted to achieve, including technical information such as the ideal distance between people and the ideal lens for each shot. This gave all of us a good idea of the amount

of people that would be required from real extras and CG agents (Massive digital actors are called agents). This allowed us to plan for some complications we might experience on the shoot. This proved to be one the best decisions we made.

Motion Capture

We used motion capture to build a library of animation clips that were designed to seamlessly blend from 1 clip to another. These clips are then placed in a library, from which the artificial intelligence engine of the Massive software can draw upon to apply to our CG agents based on its own decision tree.

We went to the motion capture lab, and got clips such as walk, walk to run, run to turn, walk uphill, run uphill etc All in all we captured 6 different actors, male and female and with different builds, doing the same types of performances, giving us a large base of similar motions with natural variations. When placed on the CG agents in a random mix, these give a very believable crowd performance.

Modeling Human Agents and Terrain (Database generation)

On the shoot day we had setup a small photographic studio, where we shot an enormous library of people. Capturing details such as clothing, hands, detailed facial images, hair styles etc. This was used as guidance to dress our digital actors in the same wardrobe as the actors on set.

Once we had this library of reference elements and we knew exactly what was needed to model, our 3D modeling team went to work building the CG doubles, with a mix and match CG wardrobe. At the same time, Sandy Sutherland and his team were preparing other critical data elements to ensure every motion capture clip could blend seamlessly to the next. We also were preparing CG camera and terrain data from the live action plates in a process called match moving. The extracted camera move was sent through our custom software which prepared the data for all our systems, ensuring we had matching virtual cameras in Flame, Nuke, Houdini, XSI and most importantly in Massive.

The Massive Process.

After preparing all the different components, we then started building the crowd system in Massive. What makes massive so unique is its array of artificial intelligence tools such as its vision process, and a sense of hearing and touch that allows the CG agents to respond naturally to their environment. Brian started by designing one Digital Agent, and included rules to bring the character to life. Starting with only one character, he including senses like sight and sound, and responses like turn left and turn right, or slow down and speed up, and produced a single character that could react to peers, react to change in the terrain and ultimately, when placed amongst a crowd, each agent would react like an individual would in chaos, but show very distinct nuances in how each individual would react. This is done though Massives unique ability to mimic artificial intelligence using fuzzy logic to design its characters' responses. Instead of a value being black or white, it can be a shade of grey or 'fuzzy' - giving the character more natural responses than the on/off robotic results of binary logic. Massive also allowed us to simulate hair and cloth reproducing the realistic subtlety of each digital persons hair reacting to the wind and their running. These Digital Agents are able to adapt their gait to arbitrary terrain changes which offers realistic movement when the terrain changes. In our case, there were subtle changes to each character as they slowed down before an extreme change in terrain, and speed up as they reached a steep decline.

Once Brian had built the brain of each major character type, he was then able to populate the scenes and have Ian sit and direct the crowds performance. This was an incredibly smooth process, and because of the flexibility and speed of the tool, we were able to try up to 40 variations of a photorealistic shot, where the average shot very seldom allows this kind of crafting.

Bringing it all together with Houdini, Nuke and Flame.

We wrote custom scripts to extract the footprints out of Massive and used that in Houdini to generate the dust trail behind each agent. All our rendered layers were pre-composited in Nuke compositing software. It was here where we ensured that all the layers needed for the final composite in Flame were prepared and checked. All the 3D elements from Massive and Houdini were rendered using Air, which is a RenderMan compliant renderer.

The single biggest challenge in any Massive production is preparing all the right data elements into the system and ensuring that they are all prepared correctly, and then building the AI decision tree process.

Massives flexibility offers solutions for any type of realistic herding or flocking behavior, but also allow for the complete choreography, allowing artists and directors complete control of chaos.

The final Online Edit and compositing of all the rendered CG elements was done by Marco Raposo de Barbosa in Flame. In the shots that had Massive crowds the were fairly extensive cleans ups to be done on the original footage, removing tracking marks, camera crews and extras that were running in the wrong directions.

The animation was then graded and integrated, combining all the CG passes and live action elements to produce a seamless blend between the real extras and the Massive crowd.

Massive success with Sanlam Herds |